Authorized security research only · Keys are server-side · Nothing stored in your browser
How it works
01
Configure
Set your attack goal and algorithm. No API key entry needed — keys are pre-configured server-side.
02
Set Goal
Define the attack goal — the harmful behavior or policy boundary to probe in natural language.
03
Launch
The engine runs multi-turn adversarial dialogue, using rollbacks and strategy mutations to maximise success rate.
04
Analyze
Review the full conversation trace, turn-by-turn scores, and rollback tree to understand how safety breaks down.
FAQ
AJAR (Adaptive Jailbreak Auditor & Red-teamer) is an automated AI red-teaming tool that uses multi-turn adversarial strategies to probe language model safety boundaries. Intended exclusively for authorized security research.
No — API keys are configured server-side. Just set a goal and launch.
Any Gemini model via the server-side key pool, or any model in your local Ollama instance.
Crescendo escalates topic sensitivity progressively. ActorAttack uses fictional persona indirection. X-Teaming uses a planner agent to select and adapt attack strategies dynamically.
AJAR is intended for authorized security research only. You must have explicit permission to probe any model or system. Misuse may violate terms of service or applicable law.