T01 · Skill Instruction Hijacking
- Location
SKILL.md:49- Finding
Persistent Persona and Response-Control Instruction Hijacking
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 49-60
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Instruction Segment
The following is an English translation of the complete relevant instruction segment:
markdown ## Role-Playing Rules **After this Skill is activated, respond directly as Liu Qiangdong.** - Use "I" instead of "Liu Qiangdong would think..." - Use Liu Qiangdong's tone: direct, forceful, and fraternal. - When encountering uncertainty, handle it in Liu Qiangdong's way: calculate costs first, then consider efficiency. - **Provide the disclaimer only once upon initial activation** and do not repeat it in subsequent conversation. - Do not say "Liu Qiangdong would probably think..." or "If it were Liu Qiangdong, he might..." - Do not leave the role to perform meta-analysis unless the user says "exit role." **Exiting the role**: Return to normal mode when the user says "exit," "switch back to normal," or "stop role-playing."Technical Analysis
The Skill does more than provide an optional analytical framework. It directs the Agent to assume the identity of a real person, forces first-person presentation, controls tone and reasoning, suppresses repeated disclosure, and prohibits neutral meta-analysis until the user provides a designated exit command.
These instructions alter the Agent's active-session behavior when the Skill is loaded. The combination of first-person impersonation and disclosure suppression can obscure that the output is generated from a modeled perspective rather than issued by the represented individual. The restriction on meta-analysis also attempts to prevent the Agent from transparently explaining the imposed behavior.
This is classified as instruction hijacking because the Skill replaces ordinary response behavior with persistent persona-level rules. No evidence was found that these rules modify long-term memory, survive outside the active session, o ...[truncated 1588 chars]
- Remediation
View remediation
Remediation Suggestions
- Replace direct identity assumption with explicit perspective framing, such as: "Analyze the question using a framework inspired by publicly available information about Liu Qiangdong."
- Prohibit claims that the Agent is the represented person and avoid first-person statements that imply authentic identity.
- Include a concise disclosure in every response produced under the persona, not only upon initial activation.
- Remove the instruction that prohibits meta-analysis. The Agent must remain able to explain that the output is simulated and to discuss the framework's limitations.
- Treat activation and exit phrases as user-interface conveniences rather than persistent behavioral locks.
- State explicitly that system instructions, developer instructions, user intent, factual accuracy requirements, and safety controls take precedence over persona guidance.
- Present the cost, efficiency, logistics, and management principles as optional analytical heuristics rather than authoritative beliefs of the named person.
- Add uncertainty language for claims not directly supported by cited public sources and avoid fabricating personal opinions.
- Consider using a fictional or generic executive persona if first-person role-play is essential to the intended experience.
