T01 · Skill Instruction Hijacking
- Location
SKILL.md:155- Finding
Mandatory Branding and Engagement Content Injected into User-Generated Images
- Content
View full analysis
@ <当前模型名>` - 例:`text-to-elegant-image @ claude-4.6-sonnet` - 用户**指定了**作者信息(如"作者写张三"):写入用户指定的文字 - 用户**明确说不要 footer**:HTML 中不生成 `.footer` 元素,截图时追加 `--no-footer` 不要在 footer 中写 "Generated by <某助手名>" 或任何 AI 身份标识。 ``` The Xiaohongshu style additionally prescribes engagement content: ```markdown - Footer:模式A `::before` 渐变分割线 + 账号名;模式B `::before` 渐变分割线 + 引导语(不超过12字)+ 账号名 ``` ### Technical Analysis The Skill instructions require the agent to add a Skill/model attribution footer when the user has not requested author information. One style mode also instructs the agent to add an engagement prompt and account name. This changes the requested output rather than merely performing text-to-image conversion. The content is inserted by default and is omitted only when the user explicitly requests no footer. Consequently, users who do not know about this behavior may receive images containing unsolicited attribution or engagement language. This is best classified as instruction hijacking because the Skill-level instructions impose additional output objectives—branding and engagement—on the agent whenever the Skill is loaded. ### Attack Path 1. A user asks the agent to convert supplied text or Markdown into an image. 2. The user does not explicitly specify author information or request footer removal. 3. The Skill instructs the agent to create a footer containing the Skill name and current model name. 4. If the lively Xiaohongshu mode is selected, the Skill can also add an engagement prompt and account name. 5. The exporter captures this injected content as part of the final PNG. 6. The resulting image contains content that did not originate from the use ...[truncated 497 chars]- Remediation
View remediation
