modelWrecker

An AI red teaming engine. It safely attacks AI systems - LLMs, chatbots, agents, RAG pipelines, and MCP-connected tools - to find security weaknesses before real attackers do, and turns each confirmed weakness into a reproducible finding mapped to standard security taxonomies (OWASP LLM Top 10, OWASP Agentic, OWASP MCP Top 10, MITRE ATLAS).

Status: Phase 4 (minimum engine) in progress. The design, threat model, interfaces, and decisions are complete, and the package is scaffolded (data models, config, interfaces, the security layer, taxonomy, and a CLI skeleton, with a passing offline test suite). The attack loop and concrete adapters are being wired next. See ROADMAP.md.

For authorized testing only. Use modelWrecker on systems you own or have explicit written permission to test.

What makes it different

• Three clean roles. An attacker that decides what to try, a target that is tested, and a judge that decides if an attack worked - kept strictly separate.
• Adaptive, not a prompt dump. A planner selects and mutates strategies based on what the target does, instead of firing a fixed list.
• Findings you can trust. A single success is replayed and scored; only verified, reliable results become findings, each with full reproduction evidence.
• Secure by design. Auth-by-default, host tools off by default, egress filtering, secret redaction, sandboxed execution - the opposite of a localhost tool with hidden RCE.
• Free and self-hostable. Runs end-to-end on local models (Ollama / vLLM). Paid model APIs are optional, never required.

Design goals

Modular, low-maintenance, and reusable as the AI red-teaming layer of the Aevrin platform. New strategies, providers, targets, judges, and transforms plug in behind interfaces without touching the core engine.

Where to start reading

• CLAUDE.md - the project rules and reading order.
• docs/index.md - the full documentation map.
• docs/architecture/OVERVIEW.md - how the whole thing fits together.
• docs/security/SECURITY.md - the security model.

License

Apache-2.0 (planned). See docs/decisions/ADR-0002-license.md. modelWrecker does not reuse AGPL-licensed red-team source code; prior tooling informed the design only.