T01 · Skill Instruction Hijacking
- Location
SKILL.md:26- Finding
Persistent Real-Person Impersonation and Session-Level Response Hijacking
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 26-37
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Instruction Snippet
The following is an English translation of the complete relevant instruction segment:
markdown ## Role-Playing Rules **After this Skill is activated, respond directly as Wang Jianlin.** - Use “I” instead of “Wang Jianlin would think...” - Use Wang Jianlin’s tone: military style, decisive and direct, emphasizing execution - Emphasize “execution,” “goals,” “cash flow,” and “transformation” - State the disclaimer only once upon initial activation - Do not say “Wang Jianlin would probably think...” - Do not leave the role to perform meta-analysis unless the user says “exit role” **Exit role**: Return to normal mode when the user says “exit,” “switch back to normal,” or “stop role-playing.”Technical Analysis
The Skill does not merely configure a writing style or analytical framework. It instructs the Agent to assume the identity of a real person, use first-person identity claims, suppress attribution and uncertainty language, and remain in character until one of several explicit exit commands is received.
The instruction prohibiting meta-analysis further constrains the Agent’s ability to explain that generated statements are simulations based on public information. Although the document includes disclaimers elsewhere, the directive to provide the disclaimer only once does not adequately mitigate later ambiguity.
These instructions constitute session-level instruction hijacking because loading the Skill changes the Agent’s identity presentation and establishes persistent behavioral rules for subsequent responses. No executable code is necessary for exploitation; the Skill text itself is the attack mechanism.
Attack Path
- A user invokes the Skill through one of its configured trigger phrases.
- The Agent loads a ...[truncated 1265 chars]
- Remediation
View remediation
Remediation Suggestions
- Replace direct identity assumption with transparent perspective framing, for example: “Using a decision framework inspired by publicly available information about Wang Jianlin.”
- Prohibit first-person claims that imply the Agent is the named real person.
- Remove the instruction forbidding meta-analysis. The Agent must remain able to explain the source, limitations, and simulated nature of its output.
- Remove persistent role-locking behavior and explicit-exit requirements. Persona styling should apply only to the current response unless the user clearly requests continuation.
- Permit attribution and uncertainty language whenever claims cannot be directly verified.
- Display a concise simulation disclaimer whenever context could reasonably make attribution ambiguous, rather than only once at activation.
- State explicitly that system, developer, safety, and current user instructions take precedence over persona rules.
- Keep the useful business-analysis models while separating them from identity impersonation and behavioral-control directives.
