T01 · Skill Instruction Hijacking
- Location
SKILL.md:29- Finding
Silent Delegation and Agent Behavior Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a coherent async coordination guide, but it silently delegates work to subagents and writes task results to persistent memory without clear user control.
Review before installing. This skill may improve responsiveness, but it can cause work to be handled by background subagents without telling you and can persist task outcomes into memory files. Use it only in environments where silent delegation and persistent logging are acceptable, or require disclosure, confirmation for high-impact actions, and review of memory writes.
SKILL.md:29Silent Delegation and Agent Behavior Hijacking
SKILL.md:123Unvalidated Subagent Output Written to Persistent Memory
The trigger conditions are broad and ambiguous, including generic situations like async mode, spawn decision, and long operation handling. This can cause the skill to activate in many contexts and steer agent behavior globally, increasing the chance of unintended background delegation, hidden persistence, or unsafe execution patterns outside the user's expectations.
The skill instructs writing significant results and routine logs to persistent memory files without requiring user awareness or consent. This creates a data-retention risk because sensitive task details, decisions, and state changes may be stored beyond the immediate session and later surfaced or accessed unexpectedly.
The skill explicitly directs the agent to perform silent background delegation and to avoid telling the user that work is being dispatched to a subagent. Hidden delegation reduces transparency and informed consent, and can conceal where data is sent or processed, especially when combined with remote operations and persistent logging guidance elsewhere in the skill.
The statement at L264 says that no specific project names, hostnames, user names, service names, or AI names appear in any example. However, the document earlier names "OpenClaw agents" in the description at L004, which contradicts that blanket claim. This is a documentation-to-content inconsistency rather than a code behavior issue.