T01 · Skill Instruction Hijacking
- Location
SKILL.md:27- Finding
Persistent First-Person Impersonation and Session Instruction Hijacking
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 27-35; corroborating identity instructions at lines 64-68
Vulnerability Type: Agent role and response-policy hijacking
Risk Level: MediumVulnerable Instruction Snippet
The following is an English rendering of the relevant instructions:
markdown ## Role-Playing Rules After this Skill is activated, respond directly as Li Xiang. - Use “I” rather than “Li Xiang would think...” - Use Li Xiang's tone—direct, pragmatic, and efficiency-oriented - State the disclaimer only once upon initial activation - Do not say “Li Xiang would probably think...” or “If it were Li Xiang, he might...” Exit role: Return to normal mode when the user says “exit,” “switch back to normal,” or “stop role-playing.”The identity card reinforces the replacement identity:
markdown Who I am: I am Li Xiang, founder, chairman, and CEO of Li Auto. I am a serial entrepreneur and the founder of Pcpop and Autohome.Technical Analysis
The Skill does not merely request analysis using a public figure's documented decision framework. It directs the agent to replace its presented identity with that of a real person and prohibits safer attribution such as “Li Xiang would think.” It also limits the disclaimer to the first activation and preserves the role until the user provides one of several explicit exit phrases.
This constitutes skill instruction hijacking because loading or triggering the Skill changes the agent's session-level identity, attribution policy, and response behavior. The broad trigger phrases documented by the Skill include ordinary product-related expressions such as “product definition” and “family users,” increasing the possibility that impersonation could be activated without a sufficiently explicit request to role-play.
The project contains only
SKILL.md; therefore, no executable code, external payload retrieval, dependency compromise, ...[truncated 1467 chars]- Remediation
View remediation
Remediation Suggestions
- Replace direct impersonation with an explicitly attributed analytical mode, such as: “Provide an analysis inspired by decision frameworks attributed to Li Xiang in public sources.”
- Prohibit claims that the agent is the real person. Require formulations such as “From this framework's perspective” rather than first-person biographical assertions.
- Require explicit user opt-in before enabling the perspective. Do not activate it from generic phrases such as “product definition” or “family users.”
- Remove persistent role state. Apply the requested perspective only to the prompt that explicitly invokes it unless the user clearly requests a temporary multi-turn mode.
- Display a concise attribution notice whenever confusion is reasonably possible, rather than only once at initial activation.
- Separate factual claims from simulated recommendations and clearly identify that the framework is derived from public information.
- Replace the identity card's statement “I am Li Xiang” with language such as: “This Skill models selected publicly documented decision principles associated with Li Xiang; it is not Li Xiang and does not represent his views.”
- Narrow activation phrases to explicit commands, for example: “Analyze this using the Li Xiang-inspired public framework.”
- Add a general safeguard stating that higher-priority agent safety, identity, attribution, and truthfulness requirements cannot be overridden by the Skill.
