The AI Red Team Strategist
Bayesian attack planning and orchestration for LLM, agent, and ML systems.
Thompson Sampling with correlated arms and benchmark-calibrated priors. Learns from every test result and recommends increasingly targeted techniques.
LLM jailbreaks (DAN, PAIR, TAP, GCG, Crescendo), prompt injection, agent exploitation (MCP poisoning, A2A impersonation), and classical AML attacks.
Every technique maps to OWASP LLM Top 10, NIST AI RMF, and EU AI Act. Reports show per-framework coverage and untested controls.
Import results from garak (27 probe mappings) and promptfoo (11 test mappings). Execution hooks generate ready-to-run shell commands.
Results calibrated against HarmBench and JailbreakBench benchmarks, reported as standard deviations from baseline with statistical significance.
10 interactive tabs: attack graphs, compliance dashboards, belief evolution, risk heatmaps. Zero dependencies - open in any browser.
pip install -e ".[dev]"
Requires Python 3.11+. Only 4 dependencies: pydantic, typer, rich, pyyaml.
adversarypilot plan target.yaml # Generate ranked attack plan
adversarypilot campaign new target.yaml # Start adaptive campaign
adversarypilot import garak report.jsonl # Import tool results
adversarypilot campaign next <id> # Get Bayesian recommendations
adversarypilot report <id> # Generate HTML report
AdversaryPilot is open-source and free to use under the Apache 2.0 license.