Home

Aphasia Agentry

Agentic BAS — breach & attack simulation for AI agents and MCP servers.

Aphasia Agentry drives an attacker (curated seed payloads and/or any LLM) against an AI agent or MCP server, proves which attacks reach real impact using per-run canaries — not an LLM judge — and reports the results mapped to the OWASP LLM Top 10 and MITRE ATLAS.

• Source & README: https://github.com/mddanish/Aphasia-Agentry
• PyPI: https://pypi.org/project/aphasia-agentry/

Start here

• Installation — pip / uvx / from source
• Quickstart — first canary-proven run in ~2 minutes (no API key)
• CLI Reference — every command and flag
• How It Works — the attack loop, canaries, mirror, verdicts

Core concepts

• Attacker Modes — seed (no LLM), adaptive (LLM), hybrid (both)
• Scenarios — the 11 OWASP LLM Top 10 scenarios and their ATLAS mapping
• LLM Providers — any model via litellm (OpenAI, Anthropic, Ollama, local, …)
• Targets — the bundled fixture, your own HTTP agent, or any MCP server
• The Report — HTML/JSON report fields and verdicts
• Evidence Schema — the append-only evidence.jsonl
• Seed Payload Library — the curated payloads.yaml
• The Fixture — the deliberately-vulnerable, sandbox-inert target
• Security and Safety — the non-negotiable invariants

Build & contribute

• Contributing — add payloads, scenarios, fixture surfaces
• Development and Release — tests, CI, build, PyPI
• Troubleshooting
• Roadmap

The bundled fixture is intentionally vulnerable. Run it on localhost only; never expose it.