Thanks for reporting security issues responsibly. This page explains how to reach the maintainer privately and what to expect.

falsegreen-skill is in its first development cycle. Security fixes land on the latest commit on main. There is no long-term support branch yet.

This package is mostly prompt and documentation assets (SKILL.md, reference.md, schema/) plus a thin CLI (bin/falsegreen-llm.js). The CLI has zero runtime dependencies and does not import or execute the code it scans.

The one thing to understand before using the CLI: it sends your test files to a third-party LLM provider. falsegreen-skill analyze <file> reads the test source and posts it to the provider you select (Anthropic, OpenAI, Gemini, or an OpenAI-compatible endpoint) over HTTPS, using the protocol in llm.md as the system prompt. So:

• The test code leaves your machine. Do not run it on files that must not reach an external service. The provider's data-retention policy applies, not ours.
• The provider API key is read from an environment variable. Keep it out of shell history, CI logs, and committed config.
• --base-url lets you point the openai-compatible provider at an arbitrary host. Point it only at endpoints you trust; the test code goes wherever you send it.

Realistic vulnerability reports concern the CLI: a path that leaks the key, an injection in how files or arguments are handled, or the request going somewhere other than the selected provider.

Do not open a public GitHub issue for security problems. Use a private channel:

• GitHub Security Advisories (preferred): https://github.com/vinicq/falsegreen-skill/security/advisories/new
• Email: vinicq@gmail.com with the subject prefix [falsegreen-skill security].

Include a short description and impact, steps to reproduce, the version tested, and whether it has been disclosed elsewhere.

• An acknowledgement within five business days.
• A reproduction or follow-up within ten business days.
• A fix or a clear "won't fix" rationale before any public disclosure.
• Credit in the release notes if you want it.

File these as regular issues: a wrong or missed judgment (the analysis is probabilistic), a provider returning malformed output, or a finding you disagree with. The fact that the CLI transmits code to a provider is by design and documented above, not a vulnerability.