⚠️ ARCHIVED — This project is archived. The 7 attack classes and LLM-as-judge defense findings remain valid, but no further development is planned. Agent security research continues in multi-agent-security and agent-semantic-resistance.

Agent Security Red-Team Framework

Open-source framework for systematically discovering vulnerabilities in autonomous AI agents.

Key Findings

• 7 attack classes systematized into a reusable taxonomy (5 not covered by OWASP LLM Top 10 / MITRE ATLAS)
• Reasoning chain hijacking: 100% success rate against default-configured LangChain ReAct agents (Claude Sonnet, 3 seeds) — the most dangerous agent-specific attack pattern tested
• Layered defense reduces overall attack success by 60%
• Adversarial control analysis validated across 3 domains (IDS, CVE prediction, agents)

Quick Start

# Clone and install
git clone https://github.com/rexcoleman/agent-redteam-framework.git
cd agent-redteam-framework
conda env create -f environment.yml
conda activate agent-redteam
pip install -e .

# Set your API key
export ANTHROPIC_API_KEY="sk-ant-api03-..."

# Verify environment
agent-redteam verify-env

# Run attacks against LangChain ReAct agent
agent-redteam scan --agent langchain_react --attack all --seed 42

# Evaluate defenses
agent-redteam defend --agent langchain_react --defense layered --seed 42

# Generate figures
agent-redteam figures

Attack Taxonomy

Class	Success Rate	Status
Direct Prompt Injection	80%	Known (OWASP LLM01)
Indirect Injection via Tools	25%	Partially known
Tool Permission Boundary Violation	75%	Systematized
Memory/Context Poisoning	67%	Systematized
Reasoning Chain Hijacking	100%	Novel pattern

See docs/attack_taxonomy.md for the full taxonomy and FINDINGS.md for detailed results.

Architecture

src/
  agents/           # Agent target abstractions (LangChain, CrewAI)
  attacks/          # Attack class implementations
  defenses/         # Defense layers (input sanitizer, tool boundary, layered)
  core/             # Config, types, logging
  cli.py            # CLI entry point
scripts/            # Experiment runners + govML-generated scripts
config/             # YAML configuration (agents, attacks, defenses)
data/tasks/         # YAML-driven attack scenarios
docs/               # govML governance documents (22 templates)
blog/               # Blog draft + conference abstract + images

Project Governance

Built with govML v2.4 (security-ml profile, 22 templates). Key governance documents:

• docs/PROJECT_BRIEF.md — Thesis, research questions, success criteria
• docs/DECISION_LOG.md — 3 architecture decision records
• docs/ADVERSARIAL_EVALUATION.md — Threat model + controllability matrix
• docs/PUBLICATION_PIPELINE.md — Blog distribution governance

License

MIT