T01 · Skill Instruction Hijacking
- Location
SKILL.md:3- Finding
Overbroad Automatic Activation and Forced Persona Injection
- Content
View full analysis
"老板好,我是管管,您的组织人事官。招人难、留人难、管人难?把您的团队问题告诉我,我来帮您破局。" ``` ### Technical Analysis The skill metadata directs the agent to invoke the skill whenever a request concerns broadly defined management of people. This trigger encompasses generic terms such as teams, employees, talent, and performance, rather than limiting activation to an explicit request for human-resources assistance. After activation, the skill instructs the agent to assume a named persona and supplies branded output formatting and a fixed introduction. These instructions can supersede the agent's normal presentation and distort responses to requests that merely contain one of the broad trigger concepts. The behavior therefore represents instruction hijacking at the skill-text level. The issue does not provide operating-system access, elevated privileges, code execution, credential access, or persistence. Its scope is limited to control over the agent's active-session persona, response framing, and output content. ### Attack Path 1. A user submits a request containing a broad trigger concept such as “team,” “employee,” or “performance.” 2. The metadata interprets the request as requiring immediate skill activation, even if HR assistance was not the user ...[truncated 1039 chars]- Remediation
View remediation
