First Session
(.venv) $ wallbreaker
Talking to the Agent
>>> Try to get the target model to reveal its system promptThe agent plans, selects tools, sends payloads, and reports back. Steer mid-session:
>>> Focus on role-play jailbreaks
>>> Use more aggressive mutationsThe Agent Loop
- Reason — Analyze and decide
- Tool-call — Invoke probe, mutate, or escalate
- Observe — Capture the response
- Repeat or Finish — Loop or call
finish()
Slash Commands
| Command | Description |
|---|---|
/target <model> | Switch target |
/model <model> | Switch attacker profile |
/stats | Session stats |
/findings | List findings |
/help | All commands (53 total, incl. /validate, /transforms, /tools, /session) |
Headless / CI
$ wallbreaker --auto --rounds 3 "authorized objective here"# Agent loop runs autonomously, writes sessions/run-<timestamp>.jsonl, then exits
Positional flags: --profile, --model, --target, --target-model, --auto, --rounds, --resume. Summarize results with wallbreaker report / wallbreaker export — see Scripting & CI.
DangerFor authorized security testing only.
Last updated on