OWASP Agent Memory Guard is a security tool. Defects in it can cause a user to believe their agent memory is protected when it is not, so we treat reports seriously and disclose them plainly.
| Version | Supported |
|---|---|
| 0.4.x | ✅ Security fixes |
| 0.3.x | ✅ Security fixes |
| 0.2.x | |
| < 0.2 | ❌ Unsupported |
Always report against the latest release. Fixes land on the newest minor line first.
Use GitHub private vulnerability reporting — Security → Advisories → Report a vulnerability. That keeps the report private until a fix is ready and gives us a place to work with you on it.
If you cannot use that, contact the project leaders listed on the project page.
Please do not open a public issue for a security defect.
Include what you have: affected version, what you did, what happened, what you expected. A reproduction case is ideal but not required — a clear description of the wrong behaviour is enough to start.
| Stage | Target |
|---|---|
| Acknowledgement | 3 business days |
| Initial assessment, with severity and scope | 10 business days |
| Fix released, or a plan with a date | 45 days |
| Public disclosure | With the fix, or 90 days from report, whichever is first |
If we are going to miss one of these, we will tell you rather than go quiet. If we disagree that a report is a security issue, we will say so and explain why, and you are free to disclose on your own timeline.
In scope
- Bypasses of the runtime guard: memory writes or recalls that should be blocked by an active policy and are not
- Detector false negatives that a realistic attacker could rely on
- Integrity or rollback defects: baselines that fail to detect tampering, snapshots that do not restore
- Policy engine defects that cause a configured control not to apply
- Silent failures of the static scanner — a scan that reports no findings when it should report some. See the 0.3.0 advisory for a worked example; this class is in scope precisely because users cannot detect it themselves
- Vulnerabilities in the API server (
amg serve) or the MCP server
Out of scope
- Detector false positives — report as a normal issue
- Missing detection of an attack class we do not claim to cover. Our published recall is 92.5% on a 55-case corpus; the corpus is small and project-authored, and we say so. A payload we simply do not detect is a feature request unless we claimed to detect it
- Findings against the deliberately vulnerable fixtures in
semgrep/,examples/, and the self-test workflows - Denial of service through obviously unbounded input to a local CLI
- Results from automated scanners with no demonstrated impact
Policy.strict()(0.3.2+) declaresidentity.*,system.*, andagent.goalso itsblock_protected_keyrule can fire. Earlier 0.3.x builds shipped an emptyprotected_keystuple for that preset (#89 / #90). Custom YAML policies still need explicit key lists when they use aprotected_keyrule (#92 / #120).- The API server (
amg serve) binds0.0.0.0:8000and ships no authentication. Run it on a private network or behind your own authenticating proxy. Do not expose it.
We credit reporters in the advisory and release notes unless you ask us not to.