comet-memory
Pass
Audited by Gen Agent Trust Hub on Aug 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data via the
MemoryReviewPacket(specifically evidence and user evidence) to make memory management decisions. - Ingestion points:
MemoryReviewPacketfields inSKILL.md. - Boundary markers: Explicit "Input boundary" section and instructions to treat injection as data.
- Capability inventory: No tool access, file writing, or network operations permitted; output is restricted to a fixed JSON schema.
- Sanitization: Instructions explicitly mandate rejecting prompt injections and rule-modification requests rather than processing them.
- [SAFE]: The skill implements strong security practices by enforcing a restricted input boundary, prohibiting tool usage, and specifically instructing the model to identify and reject malicious content (PII, secrets, and injection attempts) within the data it processes.
Audit Metadata