piigate is a safety net that keeps personal data (PII/PHI) out of an AI agent's context. A way to get protected data past it is a security issue, even when it looks like a "trick" rather than a bug.

Do not open a public issue for a bypass. Report it privately through GitHub Security Advisories ("Report a vulnerability").

Please include:

• the piigate version (piigate status) and Claude Code version (claude --version);
• the tool call or command that got through (gate bypass) or the output that leaked (filter miss);
• synthetic data only — never send real personal data, even redacted by hand.

You can expect an acknowledgement within 7 days. Fixes are released as a new tag; the advisory is published after the release.

In scope: gate bypasses (a tool call that reads a protected source without going through the filter), filter misses for formats the README lists as supported, pseudonym map or key exposure, the reverse/project approve human-confirmation checks.

Known limitations (see README, "Limitations") are not vulnerabilities: e.g. code that assembles a path at run time, nicknames, names written in lower case, data inside images. Reports that show a cheap way to close one of them are still welcome as regular issues.