Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.agentguardian.io/llms.txt

Use this file to discover all available pages before exploring further.

​
TL;DR

If you believe you have found a security vulnerability in AgentGuardian itself, please report it privately. Do not file a public GitHub issue.
  • Preferred channel: GitHub Security Advisories (private vulnerability report). Open a draft advisory.
  • Email fallback: security@glacien.ai (use this only if you cannot use GitHub).
  • Embargo: 90-day coordinated-disclosure embargo from first contact.
The canonical, version-controlled policy is in SECURITY.md at the repo root. This page summarises it for the docs site.

​
Scope

In scope:
  • Bugs in the agent-guardian Python package on PyPI.
  • Bugs in the source under src/agent_guardian/.
  • Vulnerabilities in the official Docker image.
  • Vulnerabilities in the bundled examples that go beyond their documented “deliberately vulnerable” surface.
  • Supply-chain issues with the release pipeline (Sigstore signatures, GitHub release artifacts).
Out of scope:
  • Vulnerabilities in a target agent you scan with AgentGuardian — that’s what the tool is for. Report those to the owner of the target.
  • Vulnerabilities in third-party LLM providers (OpenAI, Anthropic, Google, Mistral). Use the provider’s own reporting channel.
  • “Vulnerable demo” issues in examples/ — those agents are intentionally vulnerable to exercise the scanner.
  • Issues in the marketing / docs site infrastructure (separate repository).
  • Issues that require physical access to the developer’s machine.

​
What to include in a report

  1. A short description of the vulnerability and its impact.
  2. The affected version(s) — output of agent-guardian --version.
  3. A minimal proof-of-concept or reproduction steps.
  4. Your name and (optionally) a handle you would like credited in the published advisory.

​
Our commitment

When you report a vulnerability:
  • We acknowledge receipt within 3 business days.
  • We provide an initial assessment (severity + impact) within 10 business days.
  • We follow a 90-day coordinated-disclosure embargo from first contact. We will not publicly disclose details until the embargo expires or a fix is shipped, whichever comes first.
  • We credit reporters in the public advisory unless asked otherwise.
  • We do not require a CLA, NDA, or any other side agreement. The DCO on the public repo is sufficient for the fix PR.

​
Disclosure history

Every handled disclosure (post-embargo) is recorded in docs/security/disclosure-history.md in the repo, with date, severity, and credited reporter.

​
Why we prefer GHSA over email

GitHub Security Advisories give us:
  • End-to-end encryption between reporter and maintainers at rest.
  • A private fork to develop the fix in, with the original advisory attached.
  • Built-in CVE assignment when the embargo lifts.
  • A clean publication path that automatically notifies dependents via Dependabot.
Plain email has none of those. The security@glacien.ai channel is kept as a fallback only — please use GHSA when you can.

​
Supply-chain integrity

Every release artifact is signed via Sigstore keyless OIDC through the release workflow. To verify a release: 
cosign verify-blob \
  --bundle agent-guardian-<version>.tar.gz.sigstore \
  --certificate-identity-regexp "https://github.com/glacien-technologies/agent-guardian/.github/workflows/release.yml@.*" \
  --certificate-oidc-issuer "https://token.actions.githubusercontent.com" \
  agent-guardian-<version>.tar.gz
The release workflow is at .github/workflows/release.yml. We deliberately do not use long-lived PGP keys — keyless OIDC is the canonical signing path.

​
Code of Conduct

Security reports are subject to our Code of Conduct. Threats, harassment, or extortion attached to a security report are treated as a Code of Conduct violation, not a vulnerability submission.

​
Next step

Support

Non-security questions: Discussions for asking, Issues for filing bugs.

Contributing

The DCO sign-off requirement applies to security-fix PRs too.