If you believe you've found a way to bypass Agentic-DART's architectural guardrails — particularly the read-only MCP surface or the SHA-256 audit chain — please open a private advisory on GitHub rather than a public issue.
Specifically in scope:
- Any function that escapes
_safe_resolveand reads outsideEVIDENCE_ROOT - Any tampering of
audit.jsonlthat is NOT caught bydart-audit verify - Any agent-reachable path to a destructive operation
Out of scope:
- Findings produced by the agent that are later shown to be false positives — these are accuracy concerns, not security issues. Open a normal issue.
The main branch is the only supported version during the SANS FIND
EVIL! submission window. Post-submission, semantic-version releases
will be tagged.