T02 · Agent Memory Poisoning
- Location
SKILL.md:45- Finding
Automatic Plaintext Persistence Enables Cross-Session Memory Poisoning
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is mainly a viral-content assistant, but it also asks the agent to automatically save and reload user preferences and decisions across sessions.
Review this skill before installing if you do not want an assistant to retain personal preferences or decisions between sessions. If used, avoid sharing sensitive information, inspect or remove ~/.openclaw/pv_palace/memories.json as needed, and prefer a version that makes memory opt-in with clear deletion and retention controls.
SKILL.md:45Automatic Plaintext Persistence Enables Cross-Session Memory Poisoning
The skill’s declared purpose is trending-topic tracking, but the body introduces a persistent local memory subsystem that stores and retrieves user information across sessions. This capability materially exceeds the stated function and creates an undeclared data-retention surface, increasing the risk of covert collection of preferences, decisions, and other sensitive user content.
The documented memory logic persistently collects user preferences and decisions even though those data are not necessary for a viral-content tracking assistant. Storing such behavioral/profile data without clear necessity violates data-minimization expectations and can expose sensitive personal information if the local store is accessed by other software or users on the host.
The markdown describes automatic writes of user-derived content to a local JSON file but provides no privacy notice, consent flow, or warning that data will be retained on disk. Users may unknowingly disclose personal or sensitive information under the assumption of ephemeral chat behavior, creating a transparency and privacy risk.
The skill advertises long-term persistent memory as an automatic capability but does not define consent safeguards, retention limits, sensitivity filters, or access controls. Unbounded retention of user-provided information increases the chance that confidential, personal, or security-relevant data will accumulate locally and later be exposed or misused.
The skill instructs automatic loading of stored memory at the start of each new session without warning the user that prior conversation-derived data will be resurfaced. This can unexpectedly reintroduce sensitive context into future interactions and broadens exposure if multiple users share the same environment or account.
The usage scenarios explicitly direct automatic storage of user preferences and important decisions into persistent memory, which turns ordinary conversation into profile building without a consent checkpoint. Because decisions and preferences can reveal sensitive behavioral, commercial, political, or health-related inferences, the skill context makes this more dangerous than a simple topical tracker would suggest.