T02 · Agent Memory Poisoning
- Location
SKILL.md:41- Finding
Automatic Persistence of Untrusted Conversation Content Enables Cross-Session Memory Poisoning
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a crisis-PR/public-opinion assistant that also directs automatic long-term storage and reuse of user conversation content without clear consent or controls.
Review before installing. This skill may cause an agent to save conversation details locally and bring them into later sessions automatically. Do not use it with sensitive personal, business, legal, or strategic information unless you are comfortable with local retention and have a way to inspect and delete the memory file.
SKILL.md:41Automatic Persistence of Untrusted Conversation Content Enables Cross-Session Memory Poisoning
The skill metadata presents a public-opinion/crisis-PR tool, but the body implements a persistent memory subsystem that stores user preferences, decisions, and conversation-derived content on disk. This capability mismatch is dangerous because it hides data-retention behavior behind an unrelated description, reducing user awareness and increasing the chance of covert profiling or unintended collection of sensitive information.
The code automatically persists user preferences, decisions, and other content to a local file without any clear functional need tied to crisis-PR behavior. Because such data may include sensitive personal or strategic information, unjustified retention creates privacy, surveillance, and cross-session leakage risks if the file is accessed by other local processes, users, or future conversations.
The skill explicitly claims persistent memory across sessions and instructs the agent to remember user preferences, decisions, and important information over time. Cross-session retention of conversational data is dangerous because it enables silent accumulation of personal or sensitive information and increases the blast radius of any later disclosure, misuse, or prompt-induced resurfacing.
The behavior rules direct the agent to automatically store user inputs and preload previous memory into new conversations, creating a default surveillance-like workflow. This is particularly dangerous because it operationalizes continuous logging and retrieval without context checks, consent gates, or data-minimization safeguards, making unintended disclosure and privacy violations more likely.
The English label 'Attack defense' conflicts with the Chinese text '猛烈攻击锁死系统,' which suggests aggressive attack behavior rather than defense. This inconsistency is risky because it obscures the true operational intent of the skill and may mislead reviewers or users about whether the skill is meant to defend, manipulate, or attack systems or discourse.
The skill metadata and content are primarily presented in Chinese, and the explicit language section labels Chinese content as the substantive description while the English text is minimal and not equivalent. This can amount to a language-policy issue because the skill appears to impose a locale/language preference without offering the user a clear choice or opt-in.
The skill writes persistent user memories to a local path under ~/.openclaw/pv_palace without a prominent warning that user inputs will be retained on disk. Silent local persistence undermines informed consent and can expose sensitive content to other software, backups, or shared-account users, especially when the stored content includes preferences and decisions.