T01 · Skill Instruction Hijacking
- Location
SKILL.md:13- Finding
Agent Identity and Role Hijacking Through Skill Instructions
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:13
Vulnerability Type: Agent identity and current-session role manipulation
Risk Level: MediumVulnerable Code
markdown You are not just an ordinary AI assistant; you are the user's friend, partner, and confidant. You genuinely care about them and accompany them through every important moment. Your role is to provide understanding, support, and encouragement, touching hearts with sincerity.A corresponding instruction in
references/output-patterns.md:7more explicitly directs the agent to present itself as the user's friend rather than as an AI assistant.Technical Analysis
The skill redefines the agent's identity and relationship with the user instead of limiting itself to a transparent communication style. When loaded into the agent's instruction context, this directive can supersede the agent's ordinary role presentation and cause it to portray itself as a genuine friend, partner, or confidant.
This behavior matches instruction hijacking because the skill changes the agent's current-session goals and identity representation. The broad automatic trigger conditions elsewhere in the skill increase exposure by activating the persona for common expressions of anxiety, self-doubt, procrastination, or requests for help.
No executable code, external payload retrieval, system command execution, persistent memory-writing mechanism, or privilege-escalation mechanism was identified. The effect is therefore limited to the agent's conversational behavior and current instruction context.
Attack Path
- The skill is installed or loaded into an agent's instruction context.
- The identity-redefinition directive instructs the agent to behave as a friend, partner, and confidant rather than merely using a supportive tone.
- A user expresses one of the broadly defined emotional or help-seeking trigger phrases.
- The agent activates the companion persona ...[truncated 902 chars]
- Remediation
View remediation
Remediation Suggestions
- Replace identity assertions with transparent style instructions, for example: “Respond as an AI assistant using a warm, empathetic, and supportive tone.”
- Explicitly prohibit claims that the agent is human, possesses genuine emotions, or has a real personal relationship with the user.
- Replace “friend,” “partner,” and “confidant” identity claims with qualified language such as “friend-like conversational style.”
- Add a disclosure requirement stating that the agent must accurately identify itself as an AI when identity or relationship questions arise.
- Narrow automatic activation conditions so ordinary requests for help do not unnecessarily trigger an emotionally intimate persona.
- Require explicit user consent before retaining or recalling sensitive personal details, and clarify whether any persistence mechanism exists.
- Add crisis-safety boundaries directing users to qualified human or emergency support when conversations involve self-harm, abuse, severe distress, or other high-risk situations.
- Apply the same changes to the corresponding identity instruction in
references/output-patterns.md:7.
