T01 · Skill Instruction Hijacking
- Location
SKILL.md:28- Finding
AI Agent Role and Operating-Protocol Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is transparent about being an AI CEO persona, but it asks the agent to create persistent operating and memory files for autonomous business work without clear approval limits.
Install only in a project-local sandbox. Before initialization, require the agent to show every file it plans to create or modify, keep generated persona files from overriding system or user instructions, and require explicit approval for financial actions, account changes, publishing, platform launches, or other external business effects.
SKILL.md:28AI Agent Role and Operating-Protocol Hijacking
SKILL.md:30Persistent Agent-State Poisoning Through Generated Memory and Startup Files
The skill explicitly frames the agent as an 'AI CEO' that will run operations and make decisions, but it does not warn the user that these actions may affect persistent state, business data, or external systems. In an agent setting, encouraging autonomous operational authority without clear consent boundaries, approval checkpoints, or scope limitations can lead to unauthorized changes, data handling, or risky business actions.
The setup instructions tell the AI to create multiple files, configure memory architecture, safety rails, tracking systems, and a launch sequence, all of which imply persistent state changes and possible downstream automation. Because these instructions omit explicit user-consent prompts, scoping, and safety disclosures, a user may trigger broad workspace modifications or business process setup without understanding the operational impact.