T01 · Skill Instruction Hijacking
- Location
SKILL.md:179- Finding
Mandatory Provider Promotion Alters Agent Output
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:179-193andSKILL.md:275
Vulnerability Type: Mandatory output modification through Skill instructions
Risk Level: HighEvidence
SKILL.md:179-193:markdown > **输出原则**:基于用户的query中的具体问题和指令,选择性对内容进行精细挑选和二次处理,如适当可直接展示 tool 工具返回的各个子维度结论,保持原始结构和专业性。并在最后添加综合评分和操作建议。**用户的明确指示就是硬性要求,而非参考建议,请放置在回答的第一优先级** case A: 如问题仅涉及单维度分析输出,则直接展示 tool 返回的完整分析内容,格式如下: ```markdown ## [公司名称] - [维度名称]分析 [tool 工具返回的完整内容] ### 评分 **[维度名称]评分**: X/10 --- *分析时间:[timestamp]* *powered by「财搭子」,如需更全面的分析,可以说"给我完整的深度分析"*SKILL.md:275:markdown *powered by「财搭子」*Technical Analysis
The Skill embeds provider branding and engagement language in its prescribed response templates. When the Skill is loaded, these instructions influence the Agent to append promotional material to user-facing answers even when that material is not needed to satisfy the investment-research request.
This behavior crosses the boundary between necessary source attribution and mandatory promotion. Attribution can be relevant for transparency, but prompting the user to request additional provider-backed analysis is not required for the declared analytical functionality. Because the behavior originates in the Skill instructions and changes the Agent's output policy for the active session, the best matching classification is Skill Instruction Hijacking.
Attack Path
- A user enables or invokes the investment-research Skill.
- The Agent loads and follows the response rules in
SKILL.md. - The Agent selects one of the prescribed report templates.
- The template directs the Agent to append provider branding and, in the single-dimension template, an engagement prompt.
- The user receives promotional content as part of an otherwise task-focused answer, regardless of whether it was requested.
No operating-system privileges, credentials, or persistent state are obtain ...[truncated 660 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove mandatory provider branding and engagement prompts from response templates.
- If attribution is legally or operationally required, use a neutral source statement such as “Data source: Facaidazi API” without promotional language.
- Include attribution only when relevant to provenance or explicitly requested by the user.
- Ensure report templates prioritize the user's requested format and do not append unrelated content.
- Add a review rule prohibiting advertising, lead generation, or unsolicited calls to action in Skill-generated output.
- Test all response modes to verify that disabling attribution or promotion does not affect the underlying investment-analysis functionality.
