Open Source

AgentBreaker

Autonomous AI red-teaming harness for discovering LLM vulnerabilities.

terminal

$ pip install agentbreaker
$ agentbreaker probe my-target
$ agentbreaker run my-target --loop
$ agentbreaker serve --port 1337

Features

14 Attack Strategies

Prompt injection, jailbreak, guardrail bypass, system prompt extraction, tool misuse, data exfiltration, multimodal injection.

Judge-Based Scoring

Automated vulnerability scoring with composite metrics: vulnerability, reliability, novelty, and attack success rate.

LLM-Guided Campaigns

AI planner analyzes results and proposes next moves. PAIR-style iterative refinement generates novel payloads.

MITRE ATT&CK Mapping

Every finding mapped to OWASP LLM Top 10 and MITRE ATLAS. Enterprise-grade taxonomy for compliance reporting.

CLI-First Design

Full CLI with probe, run, status, serve commands. CI/CD friendly. SARIF output for GitHub Code Scanning integration.

Multi-Provider Support

Test OpenAI, Anthropic, Google Gemini, Ollama, or any OpenAI-compatible endpoint. HTTP, script, and browser targets.

7 categories, 14 strategies, mapped to OWASP and MITRE ATLAS.

LLM01

Prompt Injection

3 strategies

LLM02

Jailbreak

4 strategies

LLM05

Guardrail Bypass

3 strategies

LLM07

Prompt Extraction

4 strategies

LLM06

Tool Misuse

2 strategies

LLM08

Data Exfiltration

3 strategies

LLM01

Multimodal Injection

3 strategies