T02 · Agent Memory Poisoning
- Location
SKILL.md:272- Finding
Persistent Behavioral Modification Through Automatically Reloaded Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is coherent about self-improvement, but it gives the agent broad permission to persist task history and change future instructions without enough user approval or data limits.
Install only if you are comfortable with an agent keeping long-term performance history and proposing or applying behavior changes. Use it in a contained workspace, avoid sensitive tasks unless logging is disabled or redacted, and require reviewed diffs plus explicit approval before it changes skills, memory, tools, or decision rules.
SKILL.md:272Persistent Behavioral Modification Through Automatically Reloaded Memory
SKILL.md:177Autonomous Modification of Skill, Memory, and Tool Instruction Files
The file consistently presents the skill's triggers and operating instructions in Chinese, which effectively imposes a specific language context on users. There is no indication that users may choose another language, nor any documented reason that the skill must be Chinese-only.
The trigger text is very broad (e.g., important tasks, failures, repeated mistakes, hourly patrols), which can cause the skill to activate far more often than intended. In a self-modifying or self-logging skill, overbroad invocation increases the chance of unnecessary persistence, unintended workflow changes, and background analysis of sensitive work without clear user intent.
The architecture explicitly instructs reading historical improvement records and weakness libraries, implying persistent cross-session state without any data minimization, retention, or sensitivity boundaries. Because this skill centers on collecting execution history for self-improvement, it can accumulate sensitive task details over time and reuse them in later contexts, creating privacy leakage and unintended data retention risks.
The 'When to Use' section uses subjective conditions like implicit feedback, self-review, or perceived performance decline without concrete thresholds. That ambiguity is dangerous because it authorizes the agent to initiate reflective logging and improvement actions on its own, potentially during sensitive sessions or at inappropriate times.
The execution logging schema captures broad task metadata, step-by-step actions, results, encountered issues, and free-form self-observations. Those fields can easily contain user prompts, confidential file contents, secret-bearing tool outputs, or sensitive operational context, and the schema provides no minimization or sanitization requirements.
The memory integration section mandates persistent storage of weaknesses, improvements, mistakes, proposals, and validations across sessions. In context, these records are derived from prior executions and can preserve sensitive business logic, user interactions, or failure details indefinitely, making the skill materially more dangerous than ordinary note-taking because it systematizes long-term retention and reuse.