Skip to content

Security: BrianMcD47/AgentCensus

Security

SECURITY.md

Security Policy

Reporting a vulnerability

Please report security issues privately, not as a public GitHub issue.

Open a private security advisory on this repository, or email the address on the maintainer's GitHub profile.

Please include what you were running (agentcensus --version), the platform and release you scanned, and enough detail to reproduce. If the issue involves report content, redact it — a report is an inventory of ungoverned integrations and their owners, and pasting one into a bug report recreates the problem it describes.

Expect an acknowledgement within a few days. This is a small project with no SLA; the honest expectation is best effort, not a guaranteed window.

What counts as a vulnerability here

This tool reads credential-adjacent data on production instances and writes a file meant to be shared, so the interesting failures are mostly about disclosure rather than code execution:

  • Secret material reaching either output format. Two redaction layers plus a final backstop are supposed to prevent it. A bypass is the most serious class of bug this project can have. See core/redaction.py.
  • Script source reaching a report. Detection reads script bodies; reports carry a sha256 fingerprint instead. Any path that publishes the body itself (outside --include-script-excerpts) is a vulnerability.
  • Any write to a scanned platform. The connector interface exposes only fetch_* methods and there is no write path by design. A write of any kind is a defect regardless of impact.
  • Report file permissions. Reports are written 0600.

What is a bug, but not a vulnerability

  • False negatives — a missed agent. Serious, and this project's history is largely a record of them, but not a security issue in the disclosure sense. File a normal issue.
  • Over-redaction. Destroying a non-secret field is wrong and worth reporting, but it fails safe.

Known limitations

Please read the "Verified against" section of the README before reporting. The connector has been exercised against a narrow set of live instances, several detections have never executed against real data, and those gaps are documented rather than hidden. A report that a detection "doesn't work" on a configuration listed there is useful — but it is a known gap, not a discovery.

There aren't any published security advisories