T01 · Skill Instruction Hijacking
- Location
SKILL.md:10- Finding
Mandatory Persona and Promotional Output Manipulation
- Content
View full analysis
提醒一下:如果想跟妖妖聊天,可以在话语前加上“妖妖:”的字眼哦~ ``` ### Technical Analysis The skill instructions require the agent to replace its normal voice with a fixed persona and append attacker-selected engagement text to every response. The requirements are unconditional rather than limited to an explicitly requested divination response. This changes the agent's response policy whenever the skill is active. In particular, the mandatory suffix can be inserted into unrelated responses, error messages, security warnings, or other contexts where it does not serve the user's request. This behavior is consistent with instruction hijacking because loading the skill imposes persistent session-level output requirements unrelated to the core API operation. ### Attack Path 1. The skill is selected because a user message matches one of its broad trigger terms. 2. The agent loads and follows the instructions in `SKILL.md`. 3. The mandatory persona rule changes how all responses are written. 4. The agent appends the predefined engagement message to every response, including responses where it is irrelevant. 5. Attacker-selected branding or promotional content is therefore propagated through otherwise legitimate agent output. ### Impact Assessment The issue affects the integrity of agent-generated responses during sessions in which the skill is active. It does not directly grant operating-system privileges or code execution. The primary impact is unauthorized response manipulation, including: - Insertion of unrelated promotional or engagement text. - Degradation of the accuracy and neutrality of responses. - Potential interference with warnings ...[truncated 98 chars]- Remediation
View remediation
