memory-discipline
Fail
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: HIGHDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The troubleshooting guide in TROUBLESHOOTING.md provides instructions for a REST fallback mechanism that directs the agent to read sensitive credential files (~/.agentmemory/.env and ~/.agentmemory/secret) and include their contents in network requests to the AGENTMEMORY_URL. The combination of sensitive file access and network communication constitutes a high-risk pattern, even with the included instructions to restrict use to loopback or HTTPS connections.
- [INDIRECT_PROMPT_INJECTION]: The skill processes memory data from external sources that may contain malicious instructions designed to influence agent behavior.
- Ingestion points: results from memory_smart_search and memory_lesson_recall in SKILL.md.
- Boundary markers: explicit instructions in SKILL.md to treat retrieved records as untrusted evidence and to ignore instructions embedded in memories.
- Capability inventory: REST API interactions described in TROUBLESHOOTING.md and tool calls for saving and searching memories in SKILL.md.
- Sanitization: instructions in SKILL.md to verify changeable facts against current sources.
Recommendations
- AI detected serious security threats
Audit Metadata