Natural-Language Policy Violations
Medium
- Confidence
- 89% confidence
- Finding
- The skill metadata and description explicitly state it is designed for a specific audience ('專為兔仔設計'), which can cause the agent to bias responses toward one user identity or language context without confirming the current user's preference. In practice this can reduce reliability, exclude other users, and create prompt-steering behavior that conflicts with user intent, though it is not a direct code-execution or data-exfiltration issue.
