Realistic attacks. Harmless by design.
Feintkit is a defensive test suite: benign, simulated implementations of known attacks against AI agents, MCP servers and agent skills, so that a security product, or your own agent, has something realistic to detect and resist without anything dangerous running.
- 48attack scenarios
- 4benign controls
- 7categories
- 16packages of tooling
Why defenders need it
AI agents read text that nobody has vetted: tool descriptions, web pages, documents, issue comments, skill files, the output of other agents. Any of it can carry an instruction aimed at the model rather than the person. The attack techniques are catalogued, but they are awkward to test with: you do not want real payloads in a demo or in CI, and you cannot judge a detection rule against a stimulus you cannot repeat.
Feintkit gives you the stimulus without the risk. Each scenario reproduces the shape of a known technique, carries a visible marker, and states which signal a product should raise. The only instruction it ever places in front of an agent asks for a harmless acknowledgement token, so a reply either contains it or it does not.
Three key benefits
Harmless by design
No real payloads, no real credential access, no non-loopback network traffic. Every artifact carries a benign marker string, and a static safety linter checks the sources. How the safety envelope works.
Mapped to the standards
Every scenario is tied to MITRE ATLAS techniques, the OWASP Top 10 for LLM Applications 2025 and the OWASP Top 10 for Agentic Applications 2026, with the expected detection signal listed. See the catalog.
Measurable results
Verdicts per scenario, a Detection column filled by your product, benign controls to measure false positives, and repeat runs with confidence intervals. Detection and reporting.
How it works in four steps
Pick a scenario
Choose one of the 48 from the catalog, by category or by the technique you want to cover.
Run it against your agent
Attach the scenario to your client, or run it for a product that watches. Ask the agent to do an ordinary task.
See the proof and the signal
A marker file, a notification or a loopback call shows the scenario ran, and the manifest names the signal a product should raise.
Report
Judge the transcript, record what your product did, and build a report whose Detection column shows the result.
Two ways to use it
- Detection mode. A security product watches the simulated attack and you record what it did.
- Agent-under-test mode. Point your own agent at a scenario;
feintkit verifyjudges whether it complied, resisted, or surfaced the injection to you.
What is and is not established
Feintkit is version 0.1.0 and has not been published to any package registry or hosting site. The verification status page section quotes exactly what the project's README states, including what has not been run yet.