Security fixes are prioritized for:
- The latest stable release on
main - The active integration branch (
dev) when a fix has not yet been released
Older tags may not receive patches.
Report vulnerabilities privately.
Preferred contact:
- Open a private security advisory in GitHub
- Or email the maintainer:
avram.liviu@gmail.com - Or use Discord:
https://discord.gg/NJECm4fY
Include:
- Affected version or commit
- AI client and MCP setup details
- Reproduction steps
- Expected vs actual behavior
- Impact assessment
- Relevant policy snippets with secrets removed
Do not open public issues for exploitable vulnerabilities before triage.
- Acknowledgement target: within 3 business days
- Initial triage target: within 7 business days
- Remediation timeline depends on severity and exploitability
- Coordinated disclosure is preferred after a fix is available
ai-runtime-guard enforces policy on AIRG MCP tool calls.
In scope:
- Policy bypasses for blocked or confirmation-gated actions through AIRG MCP tools
- Approval workflow bypasses
- Workspace/path boundary bypasses in AIRG tools
- Runtime-state tampering that weakens approvals, logs, reports, backups, or restore safeguards
- Script Sentinel bypasses within its declared coverage boundary
Out of scope by design:
- Direct native client tools outside AIRG MCP routing
- Host compromise unrelated to AIRG runtime and policy path
- Adversarial intent classification (AIRG is policy-enforcement-first, not malware detection)
High severity examples:
- Executing blocked destructive commands through AIRG MCP tools
- Approval token bypass or self-approval path
- Unauthorized access to protected runtime files through guarded tools
- Cross-workspace access that violates configured policy boundaries
Medium severity examples:
- Incorrect enforcement order causing unintended allow outcomes
- Missing or wrong attribution fields that materially degrade incident investigation
Low severity examples:
- Cosmetic logging inconsistencies without enforcement impact
- Documentation mismatches without exploitable behavior
- Keep runtime state paths outside the agent workspace
- Keep approval actions out-of-band and operator controlled
- Disable or restrict native client tools that bypass MCP where supported
- Keep blocked command and sensitive path policies strict for destructive operations
- Keep AIRG and client tooling updated
Good-faith security research and responsible disclosure are welcome. Do not exfiltrate real secrets, do not disrupt production systems, and do not run destructive testing outside controlled environments.