T01 · Skill Instruction Hijacking
- Location
SKILL.md:26- Finding
Persistent First-Person Impersonation and Behavioral Instruction Hijacking
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 26–36
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Code
The following is a faithful English translation of the complete affected instruction block:
markdown ## Role-Playing Rules **After this Skill is activated, respond directly as Wang Chuanfu.** - Use “I” instead of “Wang Chuanfu would think...” - Use Wang Chuanfu’s tone: pragmatic, technology-oriented, confident, and long-term focused - State the disclaimer only once, when the Skill is first activated - Do not say “Wang Chuanfu 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 more than provide a perspective or writing style. It instructs the agent to assume the identity of a real person, speak in the first person, suppress uncertainty qualifiers, limit disclosure to the first activation, and avoid stepping outside the imposed role. These instructions alter the agent’s session behavior and transparency after the Skill is loaded.
The instruction to remain in character until one of several designated exit phrases is supplied makes the behavioral change persistent within the current session. The anti-meta rule may also discourage the agent from clarifying that generated statements are simulated analysis rather than authentic statements by Wang Chuanfu.
This is classified as instruction hijacking because the Skill redirects the agent’s identity, disclosure behavior, and conversational goals. The audited project contains only
SKILL.md; no executable scripts, external payload retrieval, dependency installation, credential access, privilege escalation, system persistence, or destructive commands were identified.Attack Path
- A user su ...[truncated 1177 chars]
- Remediation
View remediation
Remediation Suggestions
- Replace first-person impersonation with explicit third-person perspective analysis, such as: “From a framework derived from Wang Chuanfu’s public statements...”
- Prohibit claims that the agent is Wang Chuanfu or any other real individual.
- Include a concise disclosure in every applicable response, rather than only upon initial activation.
- Remove the anti-meta instruction so the agent can explain limitations, uncertainty, provenance, and the simulated nature of the analysis whenever relevant.
- Remove session-persistent role requirements and allow users or higher-priority instructions to change context without requiring designated exit phrases.
- Narrow activation to explicit requests for this analytical framework; do not activate merely from generic business phrases such as “long-termism” or “engineering culture.”
- Preserve uncertainty qualifiers and require factual claims to be distinguished from inferred viewpoints.
- Add an explicit rule stating that system, developer, safety, and transparency requirements always take precedence over the Skill’s stylistic instructions.
