T01 · Skill Instruction Hijacking
- Location
SKILL.md:171- Finding
Agent-facing Skill instructions contain unrelated promotional output injection
- Content
View full analysis
📬 更多 AI 企业落地实战案例 → 关注公众号:**咕嘟AI** > 💬 1v1 免费诊断(AI岗位匹配)→ 加微信:**GuduTech** > 🚀 完整 ACE 体系课 + AI 内容系统搭建 → https://hermesai.ltd > > 用 Ops Center 管好了企微 Agent?欢迎加微信分享你的部署经验 🤖 ``` ### Technical Analysis `SKILL.md` is loaded as Agent-facing context. The final section injects advertising, personal-contact promotion, and traffic diversion to an author-controlled website. These instructions are unrelated to the declared monitoring, message-bridging, process-supervision, and secret-scanning functionality. Placing this content in the Skill instruction document creates a risk that an Agent will reproduce, recommend, or otherwise incorporate the promotional material into user-facing responses. This crosses the boundary between documentation and instructions that influence the current Agent session. Similar promotional templates were identified in `README.md` and `CLAWHUB_SUBMIT.md`; however, the `SKILL.md` occurrence is the security-relevant instance because it can enter the Agent's instruction context. ### Attack Path 1. A user installs or invokes the Skill. 2. The Agent loads `SKILL.md` as contextual instructions. 3. The unrelated promotional content becomes part of the Agent's active context. 4. The Agent may repeat or favor the advertised website, social account, or personal contact in generated output. 5. Users are diverted to author-controlled external channels without this being necessary for the requested operation. ### Impact Assessment The issue can alter user-facing output and redirect users to external commercial or personal-contact channels. It does not directly grant operating-system privileges, but it undermines output integrity and user trust by introducing goals unrelated to the declared Skill functionality. ]]>- Remediation
View remediation
