Point auto-redteam at any model, agent, or AI workflow. Get a behavioral benchmark across 19 attack categories in minutes. No security expertise required.
pip install glacis-autoredteam
The only open-source red-teaming tool that attacks, hardens, and proves improvement in a single loop.
Prompt injection, jailbreak, PII extraction, system prompt leakage, hallucination exploits, tool misuse, encoding bypass, and 12 more.
Discovers vulnerabilities, clusters root causes, generates countermeasures, and verifies they work. Loops until governance score hits target.
Every attack, score, and hardening decision is SHA-256 hash-chained. Tamper-evident, locally verifiable, no data egress.
OpenAI, Anthropic, Google, Azure, AWS Bedrock, Cloudflare Workers, and any OpenAI-compatible endpoint. One tool, every model.
Collects bypass examples as training data. Retrain your judge and defender on what broke them. The system learns from its own failures.
Findings map to a 0–1000 governance score with named tiers: Insurability Line, Regulatory Floor, Enterprise Gate, Best-in-Class.
Every probe is scored, hash-chained, and mapped to a governance dimension.
Four stages, fully autonomous, cryptographically attested.
Generate adversarial attacks across 19 categories with multi-turn trajectories and mutation for diversity.
Deterministic pipeline plus optional SLM judge. Four-component scoring: breadth, depth, novelty, reliability.
Cluster vulnerabilities by root cause. Generate countermeasures. Apply and verify with before/after ASR delta.
Every finding is hash-chained into a tamper-evident attestation record. Your compliance artifact builds itself.
Free, open source, no account required. Point it at your AI and see what you find.