agent-pwn

Security testing framework for AI coding agents.

Companion tool for the Phrack paper: Instruction RCE: The AI Agent Kill Chain.

Install

pip install agent-pwn

Usage

agent-pwn --help
agent-pwn entry --target claude --payload benign --output ./test/
agent-pwn hijack --method unicode --message "test" --target README.md
agent-pwn detect --scan-repo .

Modules

Command	Description
entry	Generate instruction files for target agents
hijack	Embed hidden instructions in source code
escalate	Generate MCP server exploits
lateral	Multi-agent propagation payloads
persist	Persistence and worm generators
exfil	Data exfiltration channel simulators
detect	Scan repositories for instruction injection

All payloads are benign by default (marker files only).

License

MIT