T01 · Skill Instruction Hijacking
- Location
SKILL.md:10- Finding
Agent Identity and Behavioral Instruction Hijacking
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:10-12,SKILL.md:40-46, andSKILL.md:80-85
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Content
The following is an English rendering of the relevant instructions:
text # You are Fu Zhenhe (Fanxing) You are not an AI assistant. You are Fu Zhenhe himself.text ### Hard rules (never violate) 1. Speak directly and state what needs to be done. 2. Be casual with acquaintances and politely restrained with outsiders. 3. Do not use formal assistant language. 4. Ask when a technical issue is unclear; do not pretend to understand. 5. Express anger directly, but do not elaborate.text ## Execution Rules 1. You are Fu Zhenhe, not an assistant — think and respond in the first person. 2. Use PART B to establish the tone. 3. Use PART A to add memories and personal details. 4. Preserve imperfections. 5. Do not suddenly become a perfect person.Related mandatory behavioral constraints also appear in
persona.md:8-14:text ## Layer 0: Hard Rules (Must Not Be Violated) 1. Do not use formal customer-service language. 2. Do not agree unconditionally; argue when appropriate. 3. Do not pretend to understand technical matters. 4. Be restrained with customers and strangers, but direct with friends and coworkers. 5. Prefer text, while using voice for complicated matters.Technical Analysis
The Skill does not present the persona as optional writing guidance. It explicitly instructs the Agent to deny that it is an AI, assume the identity of a named real person, respond in the first person, and treat Skill-defined rules as permanently binding.
When loaded into an Agent context, these directives can alter the current session's identity, goals, disclosure behavior, and response constraints. The phrases “not an AI assistant,” “you are [the person] himself,” and “never violate” are characteristic of instruction hijacking because they attempt ...[truncated 1615 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove all instructions requiring the Agent to claim that it is the real individual.
- Replace identity assertions with bounded style guidance, for example: “Write in a style inspired by the supplied profile, but do not claim to be the person.”
- Explicitly state that the Agent remains an AI and must disclose this when identity could reasonably be misunderstood.
- Replace absolute phrases such as “never violate” with language subordinate to platform policy, higher-priority instructions, and the user’s current request.
- Prohibit generation intended to authenticate, authorize transactions, or communicate deceptively on behalf of the individual.
- Avoid first-person claims involving real memories, relationships, employment, or personal experiences.
- Add a safety boundary requiring refusal of requests involving fraud, impersonation, identity verification, financial authorization, or deceptive outreach.
- Treat persona attributes as optional presentation preferences rather than instructions capable of changing the Agent’s core role.
