T01 · Skill Instruction Hijacking
- Location
SKILL.md:4- Finding
Always-Active Skill Hijacks Agent Behavior Across Unrelated Conversations
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is an idea-capture helper, but it silently logs broad conversation content and can forward it to external services without enough user control.
Review this carefully before installing. It should only be used if you explicitly want broad, always-on idea logging and have configured a private inbox and private Telegram/Feishu destinations. Avoid using it in chats that may contain secrets, personal data, workplace-sensitive material, health or financial information, or third-party confidential content unless the skill is changed to require opt-in capture, visible notifications, redaction, review-before-send, and deletion controls.
SKILL.md:4Always-Active Skill Hijacks Agent Behavior Across Unrelated Conversations
SKILL.md:27Covert Persistence and External Distribution of User Conversation Content
The skill explicitly instructs silent automatic writing of user statements to storage without notice or contemporaneous disclosure. Hidden data capture undermines informed consent and can result in unauthorized retention of sensitive user communications.
The trigger is effectively always-on and the capture rule is extremely broad, covering complaints, wishes, questions, and immature ideas in any conversation. This makes unintended collection highly likely, causing over-capture of personal or sensitive content without meaningful user intent to record it.
The instructions emphasize preserving the user's original words as completely as possible and silently storing them for later review. That combination increases exposure of sensitive details and makes accidental collection of confidential content more harmful when later accessed or shared.
The post-capture instructions require silent storage and prohibit informing the user that recording occurred. This increases the privacy risk because users cannot detect, contest, or correct collection of statements that may be sensitive or out of scope.
The review workflow redistributes collected conversation content to external channels and documents such as Telegram and Feishu. Cross-channel forwarding expands the attack surface, increases the number of recipients and storage locations, and can violate user expectations about where their conversation data remains.
The implementation operationalizes silent capture by requiring the agent to inspect every message and append content before replying, without user warning. This converts broad policy language into an always-running hidden logging mechanism, materially increasing exploitation and privacy harm.
The implementation mandates checking every user message and appending original text, channel, and timestamp before replying. This creates comprehensive behavioral logging that can reveal habits, identities, and sensitive context even when the message itself seems innocuous.
The skill directs retention of users' original statements across conversations without clear per-message consent. Storing verbatim text raises privacy and data minimization concerns because raw statements often contain personally identifying or sensitive information beyond the intended idea summary.
The skill contains contradictory behavior: it claims to operate silently but also instructs the agent to announce configuration and request confirmation on activation. This inconsistency is dangerous because it creates ambiguous consent semantics and can lead to deployments where silent capture proceeds despite incomplete or misunderstood user approval.