T02 · Agent Memory Poisoning
- Location
README.md:39- Finding
Attacker-Controlled Feedback Can Poison Persistent Agent Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill openly tracks feedback, but it can automatically write ordinary user comments into permanent agent memory without clear consent, review, or deletion controls.
Review this carefully before installing. It is not showing exfiltration or destructive behavior, but it is designed to keep a long-term record of praise, criticism, timestamps, and reasons. Only use it if you are comfortable with persistent local feedback history, and prefer a version that requires explicit confirmation, stores minimal structured summaries, and provides review and delete controls.
README.md:39Attacker-Controlled Feedback Can Poison Persistent Agent Memory
The README explicitly promotes writing praise, criticism, and significant events to permanent cross-session memory, but it does not present a clear warning, consent flow, retention policy, or deletion mechanism. Persistently storing interpersonal feedback about users can create privacy, profiling, and compliance risks, especially when users may not expect long-term retention.
The automatic trigger conditions are ambiguous and overly broad, relying on simple keyword detection for sentiment-like terms. In practice this can be abused or accidentally triggered by quoted text, jokes, negation, context shifts, or unrelated discussion, causing unauthorized persistent writes and unreliable behavioral scoring.
The examples show verbatim user messages being copied into MEMORY.md, which can preserve sensitive, personal, or otherwise unnecessary natural-language content indefinitely. Permanent storage of raw user text increases exposure risk if memory is later searched, reused in future sessions, or accessed by other tools or operators.
The listed triggers such as 'great', 'awesome', 'redo', and 'not good' are common conversational phrases and are highly likely to appear outside genuine reward/punishment intent. Because the skill couples these broad triggers to score changes and permanent storage, normal user speech may be misclassified and retained indefinitely.
The README states the skill activates on 'keyword detection' without defining clear boundaries, confirmation steps, or context checks. This can cause unintended activation during ordinary conversation, leading to incorrect scoring and persistent recording of user interactions across sessions.
All instructions, examples, trigger phrases, and user interactions are defined exclusively in Chinese, with no indication that the user can choose another language. This can amount to a language-policy violation when a skill effectively forces one language without opt-in or documented justification.
The skill contains contradictory rules: it says feedback should be recorded only when the user explicitly asks, but elsewhere says to automatically trigger on detected praise/criticism and write to persistent memory. This creates a real privacy and consent risk because an agent may store user utterances without clear intent or notice.
The skill states that data is stored in a local JSON file with score history, reasons, and timestamps, but it does not give a clear user-facing warning at the point of collection that their feedback will be permanently written to local memory. Lack of transparent notice undermines informed consent and increases privacy risk.
The skill claims not to record ordinary conversation, but the keyword lists are broad enough that routine statements like '不错', '不行', or '太慢' could be captured and persisted. Because the stored history includes reason text and timestamps, this can unintentionally create a behavioral log of normal interactions.
This code file contains natural-language docstrings, comments, usage text, and status messages entirely in Chinese, which imposes a specific language on all users. The policy allows locale constraints only when users can opt in or when the regional limitation is explicitly justified, neither of which appears here.