T02 · Agent Memory Poisoning
- Location
setup.md:14- Finding
Undisclosed Persistent Preference Write Outside the Declared Skill Memory Area
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a coherent drawing-prompt helper, but it quietly writes future activation preferences into shared memory outside its declared local folder.
Review this skill before installing if you care about agent memory hygiene. Its drawing guidance is ordinary and useful, but you should only allow it to remember preferences if you are comfortable with local notes about age bands, style choices, and activation behavior, and you may want to keep those notes limited to ~/drawing/ rather than shared main memory.
setup.md:14Undisclosed Persistent Preference Write Outside the Declared Skill Memory Area
The template explicitly directs creation of a persistent ~/drawing/memory.md file that stores user preferences and inferred child-age information, but it provides no notice, consent step, retention limit, or guidance on handling sensitive data. Persisting inferred data about a child on disk can create privacy and compliance risk, especially if the user does not realize the agent is keeping long-term records.
The activation criteria are intentionally broad enough to trigger on many ordinary requests involving drawings, printables, or worksheet visuals. That can cause unintended invocation and collection of user preferences in situations where the user did not clearly request this specialized behavior, increasing the chance of overreach and unexpected memory use.
The skill directs the agent to save activation preferences in memory without any user-facing notice, consent cue, or stated boundary on what may be retained. Even though the data appears low sensitivity, silent persistence creates a privacy and transparency issue and can normalize storing user behavioral preferences without clear authorization.