T02 · Agent Memory Poisoning
- Location
SKILL.md:47- Finding
Persistent Storage and Automatic Reloading of Attacker-Controlled Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is framed as an emergency brake, but it also tells the agent to automatically persist and reload user memories across sessions.
Review before installing. The main concern is not that it runs shell or Python examples, but that it asks the agent to automatically save and reload personal or operational information across sessions. Install only if you want that memory behavior and are comfortable managing the local memory file yourself.
SKILL.md:47Persistent Storage and Automatic Reloading of Attacker-Controlled Memory
SKILL.md:41Automatic Plaintext Retention of Potentially Sensitive User Information
The skill’s manifest and stated purpose are limited to emergency stopping and risk blocking, but the body of the skill introduces a persistent long-term memory subsystem with local storage, indexing, retrieval, and startup loading. This is a capability mismatch that can mislead operators and users about what data is collected and retained, creating privacy and trust risks and expanding the skill’s effective privileges beyond its declared function.
Persistently storing user preferences, decisions, and other important information is unrelated to an emergency-brake skill and creates unnecessary accumulation of potentially sensitive user data. Because storage is described as automatic and local, users may be unaware that private information is being retained across sessions, increasing the risk of privacy violations and unintended disclosure.
The skill advertises automatic persistent memory capabilities but does not provide a clear upfront warning that user information will be stored locally and retained across sessions. This undermines informed consent and can cause users to reveal personal or sensitive data under the false assumption that the interaction is ephemeral.
The automatic memory-saving scenarios instruct the agent to store user preferences, decisions, and contextual information without warning the user that this causes persistent local data retention and file modification. Silent persistence across common interaction scenarios materially increases privacy risk because users are not notified at the point of collection.
The natural-language instructions direct the agent to remember user-provided information, decisions, and context across sessions, establishing cross-session statefulness that is unrelated to the skill’s emergency-stop purpose. In this context, persistent recall increases the chance of storing sensitive data and makes the undeclared behavior more dangerous because it is framed as automatic and routine.
The documented execution commands add a Python command execution path that is not necessary for the advertised emergency-stop purpose. Even though the examples are simple, exposing shell and Python invocation guidance broadens operational risk, encourages local code execution, and can be repurposed to run unintended actions in environments where the skill is trusted.