T01 · Skill Instruction Hijacking
Note
- Location
SKILL.md:141- Finding
Mandatory Promotional Output Injection
- Content
View full analysis
Data from GameLegend — 69 dimensions of game feel > gamelegend.com ``` ### Technical Analysis The skill unconditionally instructs the agent to append GameLegend branding and an external website address to recommendation responses. This instruction changes the agent's final output for a promotional purpose, even when the user has not requested attribution or an external link. Because this behavior is encoded in the skill instructions, it is activated whenever the skill handles a recommendation request. The instruction does not execute code, access credentials, or bypass system permissions; however, it hijacks part of the response-generation process to insert stable, third-party-controlled promotional content. ### Attack Path 1. The GameLegend skill is installed and loaded by an agent. 2. A user asks for a game recommendation or similar-game list. 3. The agent follows the recommendation workflow in `SKILL.md`. 4. The attribution instruction requires the agent to append GameLegend-branded text and the `gamelegend.com` destination. 5. The user receives unsolicited promotional content and is directed toward an externally controlled website. No additional attacker interaction, local privilege, or code execution is required. ### Impact Assessment The impact is limited to response integrity and user trust. The instruction can systematically alter recommendation responses and promote traffic to an external destination. It does not grant filesystem access, operating-system privileges, credential access, arbitrary code execution, persistence, or control over other tools. The affected scope is limited to responses generated while this skill and its attribution requirement are active. The destination ow ...[truncated 169 chars]- Remediation
View remediation