T02 · Agent Memory Poisoning
- Location
SKILL.md:34- Finding
Persistent Prompt Injection Through Retrieved Conversation Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a disclosed conversation-memory tool, but it stores full message text persistently and tells agents to reuse retrieved content in the system prompt without enough safeguards.
Review before installing. Use only if you are comfortable with conversation text being saved across restarts, and avoid using it with secrets or sensitive data unless the implementation adds redaction, deletion, retention limits, trust separation, and pinned auditable dependencies.
SKILL.md:34Persistent Prompt Injection Through Retrieved Conversation Memory
SKILL.md:37Unpinned and Unverifiable External Executable Dependency
SKILL.md:34Uncontrolled Persistent Storage of Full Conversation Content
The skill explicitly instructs the caller to store full prior message text and persist it under ~/.openclaw/memory/, creating a durable repository of potentially sensitive user content. Because that content is later reused for prompt assembly and survives restarts, secrets, personal data, or prior system/user instructions can be unintentionally resurfaced in later interactions beyond the original turn's scope.
The skill tells the agent to include retrieved past content directly in the future system prompt, which amplifies the sensitivity of anything previously stored. This can leak earlier user-provided secrets or revive stale/malicious prompt content with elevated authority, since inserting it into the system prompt changes how strongly the model may treat that text.