Email: open a GitHub security advisory (preferred) or contact the maintainer directly. Do not open a public issue for exploitable flaws.
Include: description, impact, reproduction (the AttackBench scenario format is
ideal — intentguard/attackbench/schema.py), affected versions/commits.
In scope: the IntentGuard backend (backend/intentguard/), its API, the
dashboard, the Docker packaging, and the security claims made in
docs/SECURITY_MODEL.md / docs/THREAT_MODEL.md (a documented limit that fails
in practice is a vulnerability).
Out of scope: the mock tools' sandbox content (attacker pages under
*.example are intentional), social engineering, physical attacks, and
vulnerabilities in undeployed research code paths.
- Acknowledgement: 72 hours
- Triage & severity: 7 days
- Fix or mitigation for criticals: 30 days
If you test IntentGuard: only against your own deployment, only the sandbox mock tools, never real third-party services. The AttackBench harness exists precisely so adversarial testing never needs to leave the sandbox.