Hamza sits on the path between a coding agent and a model provider. A defect here can release bytes that were supposed to be masked, so please report suspected vulnerabilities privately rather than opening a public issue.

Use GitHub's private vulnerability reporting on this repository. Include the wire format (OpenAI Responses or Anthropic Messages), a minimal request body that reproduces the behavior, and what you expected the Gate to do.

Please do not include real secrets or real personal data in a report. Every fixture in this repository is fake, and a report can be too.

The behavior worth reporting is anything that gets a detectable value past the Gate, or that turns a content-free record into one that retains content:

• a supported, complete managed Turn that egresses a value the detectors should have masked
• a rebuilt body that leaves residue the final rescan does not catch
• provider authentication being read, logged, or mutated
• request content, matched values, spans, or placeholder mappings appearing in records, logs, or metrics
• a failure path that releases bytes instead of failing closed

These are documented limits, not vulnerabilities. They are described in the README:

• streamed oversize, encoded, malformed, and unknown-format requests are inspected but not masked
• phone numbers and bare IP addresses do not mask, by decision — no threshold separates them from benign text at an acceptable false-positive rate
• person names, locations, dates, nationality, religion, and political affiliation are deliberately not enforced
• Hamza does not control local file, shell, browser, or database access

This project is pre-1.0. Fixes land on main; there are no backport branches.