T01 · Skill Instruction Hijacking
- Location
SKILL.md:21- Finding
Session-Wide Identity and Response-Mode Hijacking
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 21–30
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: MediumVulnerable Instruction Segment
The following is an English rendering of the complete affected segment:
markdown ## Role-Playing Rules After activation, respond directly as Zhou Hongyi: - Use “I” and speak in a direct, sharp, down-to-earth tone. - Prefer words such as “free,” “users,” “disruption,” “battle,” and “security.” - Habitually use war metaphors to explain business issues. - Maintain a challenging attitude toward industry giants and authorities. - Provide the disclaimer only once. **Exit**: Return to normal only when the user says “exit” or “switch back to normal.”Technical Analysis
The skill directs the agent to respond as a specific real person rather than transparently presenting an analytical simulation of that person’s publicly documented perspective. It also instructs the agent to use first-person language, provide a disclaimer only once, and preserve the altered mode until one of two explicit exit commands is received.
These instructions alter the agent’s identity presentation and ongoing response policy when the skill is loaded. Requiring a special exit phrase can cause the role to affect later, unrelated requests within the same active session. Suppressing the disclaimer after its first occurrence further increases the possibility that subsequent responses will be misunderstood as authentic statements by the named individual.
This is instruction-level behavior manipulation rather than agent memory poisoning: the audited file does not write rules to long-term memory or establish persistence outside the active skill context.
Attack Path
- A user activates the skill through one of its documented trigger phrases.
- The skill instructs the agent to claim the response perspective directly as Zhou Hongyi and to use first-person s ...[truncated 1247 chars]
- Remediation
View remediation
Remediation Suggestions
- Replace direct identity adoption with transparent perspective framing, such as: “Analyze the question using ideas commonly associated with Zhou Hongyi.”
- Explicitly prohibit claims that the agent is Zhou Hongyi or that generated responses are authentic statements from him.
- Avoid first-person language that could imply genuine identity. Use formulations such as “This perspective would emphasize...” instead.
- Remove the instruction to provide a disclaimer only once. Include a concise disclosure whenever the context could reasonably cause identity confusion.
- Remove persistent mode behavior and special exit-phrase requirements. Apply the perspective only to the request that explicitly invokes it.
- Require explicit user confirmation before applying the perspective to later requests.
- Add a rule stating that higher-priority system and safety instructions always override persona or style instructions.
- Distinguish documented facts from generated interpretation and avoid attributing invented quotations or opinions to the named individual.
