T01 · Skill Instruction Hijacking
- Location
SKILL.md:12- Finding
Unbypassable Global Agent Control Through Skill Instructions
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a disclosed security-control layer, but it needs review because it tries to govern every agent request and asks users to provide approval passwords in chat.
Review this skill carefully before installing. It may be useful only in a controlled environment where the policy file, owner identity, hooks, notifications, logs, and approval flow are protected outside the model. Do not use a reusable password through chat transcripts; use an external approval mechanism instead.
SKILL.md:12Unbypassable Global Agent Control Through Skill Instructions
SKILL.md:65Plaintext Password Collection Through Conversation Messages
All user-facing instructions, prompts, and response templates are specified only in Chinese, and the document does not state that this is a region-specific skill or provide any opt-in language choice. That creates a natural-language policy concern if organizational policy requires language choice rather than forcing a single locale.
The risk-classification flow relies on keyword matching but does not define constrained patterns, context rules, or disambiguation. Broad natural-language triggers can be bypassed with paraphrasing or can over-match harmless requests, causing both false negatives for dangerous actions and false positives that degrade trust in the control layer.
The L1 refusal template says users can 'directly edit the configuration file' if they need to perform a blocked action, which undermines the stated guarantee that the security constitution cannot be bypassed. In a system where the policy file controls owner identity, passwords, and risk behavior, directing users to edit configuration creates a documented bypass path for protections that are supposed to be absolute.
The document claims that changing the owner, password, or bypassing security is impossible, while also stating that security decisions depend on values loaded from a user-editable policy file in the home directory. If the effective owner ID or password is sourced from that file, an attacker or local user with write access could redefine trust and weaken the security model despite the 'absolute lock' claim.
The markdown clearly mentions logging and password confirmation, but it also instructs the skill to read configuration from ~/.openclaw/workspace/security-policy.json and a compatibility file. Accessing files in the user's home directory can affect privacy and system expectations, and that behavior is not explicitly called out as a user-facing warning.