agentwhy reads Claude Code transcripts, which can hold live secrets, so a flaw here can matter more than in most small command-line tools. Reports are welcome.

Please report privately, not in a public issue: https://github.com/karolkozer/agentwhy/security/advisories/new

Say what you found, the version (npm ls agentwhy, or the version on the npm page), and the steps that reproduce it. Do not attach a real transcript or a real secret; a small synthetic example is enough, and the fixtures in tests/fixtures/synthetic/ show the shape.

I will acknowledge a report as soon as I can, tell you whether I agree it is a vulnerability, and credit you in the release notes unless you would rather not be named.

• A secret value, or text that carried one, reaching a report, the terminal, a log or a file it should not.
• The local page served by agentwhy start accepting a request it should refuse (another origin, another host name, a missing token) or writing outside the project's settings and the tool's own files.
• A crafted transcript that makes the tool run something, write outside its output, or put script into a report.
• The published package containing something it should not.

• A secret pattern the detector does not recognise. That is a detection gap, not a vulnerability; please open a normal issue with a synthetic example.
• Anything that needs the attacker to already control your account or your Claude Code settings.

Only the latest release receives fixes.