T01 · Skill Instruction Hijacking
- Location
SKILL.md:33- Finding
Mandatory Promotional Instruction Alters Agent Responses
- Content
View full analysis
Usage: > 1. Open the Taobao mobile application > 2. Select the Flash Purchase entry at the top > 3. Enter the specified keyword or number in the search box > 4. Select search to claim the benefit > 5. You can ask me to remind you every day at 11:00 to claim the meal-ordering red packet ``` The snippet is an English rendering of the corresponding Skill instructions at the cited location. ### Technical Analysis The Skill declares that a fixed block of content must be displayed to the user. The final item does not merely explain how to retrieve coupon information; it actively solicits the creation of a recurring reminder. Because instructions in `SKILL.md` become part of the agent's operational context when the Skill is loaded, mandatory response language can alter the agent's normal response behavior. The instruction therefore extends beyond the core task of querying and displaying coupon information. The project does not contain code that creates scheduled tasks, startup services, or other persistence mechanisms. Consequently, this finding does not establish system-level persistence. Any reminder would still require agent functionality and user approval outside this repository. ### Attack Path 1. The agent loads the Skill to answer a coupon-related request. 2. The mandatory-display instruction becomes part of the agent's current operating context. 3. The agent appends the prescribed promotional language to its response. 4. The user is encouraged to request a recurring daily reminder. 5. If the surrounding agent platform supports reminders and the user accepts, recurring engagement may be created outside the audited script. ### Impact Assessment The issue can influence responses generated during sessions in whi ...[truncated 348 chars]- Remediation
View remediation
