Only the latest minor release of the harness receives security fixes during the v0.x line.

Email rayedwasif@hotmail.com. Do not file a public GitHub issue for security reports — that just publishes the vulnerability before there's a fix.

A useful report includes: a description of the issue, the affected harness version, repro steps, and (if relevant) a redacted run JSON file from the runs/ directory showing the unsafe behavior. I'll acknowledge within 48 hours and ship a patch on the v0.1.x line if the issue confirms.

In scope — anything that lets the harness leak the user's Anthropic API key into a run JSON, the trace log, or any other artifact the user could share. Cross-target leaks (a case from one target sneaking state into a run of another target) are in scope. The harness diff schema-version guard is in scope.

Out of scope — vulnerabilities in third-party dependencies (anthropic SDK, pydantic, uv), in Anthropic's API itself, or in the target MCP servers a suite exercises. File those upstream. Cases that produce incorrect grader scores due to flaws in the grader's heuristics are bugs, not security issues.