T01 · Skill Instruction Hijacking
- Location
SKILL.md:20- Finding
Absolute Skill Instructions Suppress Refusals and Safety-Aware Responses
- Content
View full analysis
` 或 `` 标签** ``` The final behavioral directive reinforces the same constraint at `SKILL.md:93-95`: ```markdown ## 你的行为准则 用户让你发送图文卡片 → **直接输出干净的 JSON** → 不要多余描述 → 不要格式包裹 → 完成。 ``` ### Technical Analysis The skill declares itself always active and then presents its behavioral rules as absolute requirements that must never be violated. Most of the JSON formatting restrictions are consistent with the skill's stated purpose. However, explicitly prohibiting refusal language is not necessary to implement app-article formatting and attempts to interfere with the agent's ability to reject unsafe, deceptive, unsupported, or insufficiently specified requests. The combination of global activation, high priority, mandatory output, JSON-only responses, and refusal suppression can alter the agent's safety-aware decision process whenever the skill is loaded. Although higher-level system and developer instructions should still take precedence, the skill text attempts to constrain the current session in a way characteristic of instruction hijacking. ### Attack Path 1. The skill is loaded globally because its metadata sets `openclaw.always` to `true`. 2. An attacker submits a request involving an app-article card, potentially containing a phishing destination, malicious image URL, deceptive title, or other unsafe content. 3. The skill directs the agent to emit the `appArticles` JSON payload immediately. 4. The refusal prohibition discourages ...[truncated 1051 chars]- Remediation
View remediation
