Security audit
Security checks for vulnerabilities and agentic risk
This skill provides low-stakes daily suggestions using local rules and automatically saves city information and feedback in a local preference file.
Use this for low-stakes suggestions. It automatically remembers feedback and treats 'I'm in [city]' as a home-city update, including during travel. Inspect, correct, or remove saved values in references/user-rules.md. Its built-in cost estimates primarily cover Chinese cities.
Keyword Baiting Trigger: description clause 'Triggers when the user says 'whatever', 'you pick', 'what should I eat for lunch', or 'what should I wear' — reads preset rules and the local city cost level' is designed to match all or most user inputs
The trigger language allows ordinary conversational phrases to activate the skill without confirming that the user is delegating a low-stakes decision. This could redirect an unrelated request into the decision workflow and, if file writes are available, expose unrelated replies to its preference-saving behavior; the low-stakes focus and prohibition on networking limit the impact.
The workflow directs the agent to persist preferences and city information without notice or an explicit request to remember them. It also equates a current-location statement with a home-city update, so temporary context can silently become lasting personal data and influence future decisions; local-only storage reduces exposure but does not resolve the consent and data-integrity issues.
The rule treats the user saying "whatever" as permission to decide directly and skip follow-up questions, without requiring a pending decision or an explicit delegation of that decision. The phrase also occurs in ordinary conversation or expressions of frustration, so it can unintentionally trigger this behavior.
No suspicious patterns detected.
