remember
Pass
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: SAFECREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [CREDENTIALS_UNSAFE]: The TROUBLESHOOTING.md file contains explicit instructions for the agent to access sensitive credential files, specifically
~/.agentmemory/secretand~/.agentmemory/.env, to facilitate a REST fallback if standard tools are unavailable. While these are vendor-specific resources, instructing an agent to manually read its own secrets increases the risk of accidental exposure or exploitation. - [INDIRECT_PROMPT_INJECTION]: The skill creates a significant surface for indirect prompt injection by ingesting raw user input through the
$ARGUMENTSvariable and storing it directly into a long-term database via thememory_savetool without any sanitization. - Ingestion points: The
$ARGUMENTSvariable in SKILL.md captures raw user input. - Boundary markers: There are no boundary markers or delimiters used when passing content to the
memory_savetool. - Capability inventory: The
memory_savetool writes to a persistent storage backend that is likely retrieved by a correspondingrecallskill. - Sanitization: No filtering, escaping, or validation of user-provided content is performed, allowing potentially malicious instructions to be stored and later executed during memory retrieval.
Audit Metadata