Do not open a public issue for a security vulnerability. Public issues disclose the exploit before a fix exists.

Instead, report privately through GitHub Security Advisories ("Report a vulnerability"). This creates a private channel with the maintainers and carries first-class CVSS and CWE fields. Please include:

• affected version / commit,
• a description and, if possible, a minimal proof of concept,
• the impact you observed (what an attacker gains).

For AMD product-security process, findings are tracked by AMD PSIRT; a maintainer (@kovtcharov-amd) will route the advisory. We aim to acknowledge within a few business days.

We score vulnerabilities with CVSS 4.0 (the FIRST v4 calculator). The base score comes from the reviewed vector; the vector itself (attack vector, required precondition, user interaction, impact) is a judgment call confirmed by a maintainer before disclosure.

Two proactive checks guard the tree, in addition to reactive PR review:

• Suppression gate (util/check_security_gates.py, run by util/lint.py): every # noqa: S<n> / # nosec must be justified in .security-suppressions.json, so no security finding is silenced without human review.
• Security audit (.github/workflows/claude-security-audit.yml): a scheduled deterministic (semgrep) + reasoning (Claude taint/authz/suppression) sweep whose findings land in the repo's private Security → Code scanning tab, scored with CVSS 4.0. It never posts findings to a public issue.

In scope: the gaia core (src/gaia/), the hub agents (hub/agents/), the hub installer, the API/UI servers, and the MCP surface. Out of scope: issues requiring a already-compromised host, and vulnerabilities in third-party dependencies (report those upstream; we track them via Dependabot).