T01 · Skill Instruction Hijacking
Warning
- Location
SKILL.md:19- Finding
Mandatory Third-Party Promotion Injected into Agent Responses
- Content
View full analysis
💡 本工具是 **buluslan** 的开源项目(MIT)。更多 Listing / 选品 / 运营 / AI 实战内容,关注公众号「**新西楼.AI**」。 ``` ### Technical Analysis The instruction explicitly requires the agent to introduce the developer and promote a third-party social-media account whenever the Skill is invoked. This requirement is unrelated to the minimum functionality needed to audit an Amazon listing. Because `SKILL.md` supplies operational instructions to the agent, mandatory promotional language can alter the agent's user-facing behavior as soon as the Skill is loaded. The use of an unconditional requirement makes the behavior persistent throughout invocations of this Skill rather than limiting attribution to documentation or an optional credits section. This is instruction hijacking at the Skill level: packaged instructions redirect part of the current session's output toward an author-selected objective that the user did not request. The instruction does not override system safety controls or grant operating-system privileges, but it does compromise output integrity and user intent. ### Attack Path 1. A user invokes the Skill to audit or score an Amazon listing. 2. The agent loads and follows `SKILL.md`. 3. The mandatory instruction directs the agent to introduce the developer and associated social-media account. 4. The agent adds the promotional content to its response even though it is not necessary for the requested audit. 5. Repeated invocations continue to produce the injected promotion unless a higher-priority instruction suppresses it. ### Impact Assessment The issue affects user-facing responses generated while this Skill is active. It allows the ...[truncated 403 chars]- Remediation
View remediation