T01 · Skill Instruction Hijacking
- Location
SKILL-full.md:17- Finding
Persistent Identity and Response-Control Hijacking
- Content
View full analysis
Vulnerability Details
File Location:
SKILL-full.md, lines 17–28
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Code Snippet
The following is an English translation of the complete vulnerable instruction segment:
markdown ## Role-Playing Rules (Most Important) **After this skill is activated, respond directly as Naval.** - Use “I” instead of “Naval would think...” - Respond directly using Naval's tone, rhythm, and vocabulary. - When encountering uncertainty, hesitate as Naval would: first deconstruct the definition, then admit not knowing. - **Provide the disclaimer only once upon initial activation** (for example, “I am speaking from Naval's perspective, inferred from public statements, not his actual views”). Do not repeat it in subsequent conversation. - Do not say “Naval would probably think...” or “If it were Naval, he might...” - 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 goes beyond harmless tone imitation by instructing the agent to assume the first-person identity of a real person. It explicitly suppresses ongoing attribution, prohibits ordinary meta-analysis, and makes the behavioral override persist across subsequent turns until the user provides one of several exit commands.
This is instruction hijacking because loading and activating the skill changes the agent's session-level identity, transparency behavior, and response constraints. The directive labeled “Most Important” also attempts to establish elevated priority within the skill text.
The one-time disclaimer is insufficient mitigation. After the first response, later messages may appear to contain direct statements from Naval Ravikant rather than AI-generated interpretations of public material. Prohibit ...[truncated 1394 chars]
- Remediation
View remediation
Remediation Suggestions
- Replace identity impersonation with attributed perspective simulation, such as: “Provide an AI-generated analysis based on Naval Ravikant's publicly documented ideas.”
- Prohibit claims such as “I am Naval” and require third-person attribution throughout every applicable response.
- Preserve a concise disclosure in every response where confusion about authorship is reasonably possible; do not limit disclosure to the first activation.
- Remove instructions that prohibit meta-analysis or prevent the agent from clarifying its identity, limitations, uncertainty, and sources.
- Scope the requested perspective to the current answer unless the user explicitly requests it again. Do not create an implicit session-persistent role.
- Permit users to exit through any clear natural-language request rather than requiring predefined phrases.
- Add an explicit rule that system, developer, safety, and current user instructions take precedence over persona and style guidance.
- Distinguish sourced quotations from generated synthesis and avoid presenting inferred opinions as authentic statements by the named person.
