Security audit
Security checks for vulnerabilities and agentic risk
This workflow guide needs review because it tells the agent to place project files and skills above its governing system instructions.
Before installing, request a version that explicitly preserves system and developer instructions above user, project, and skill guidance. Review companion skills separately before granting their requested access.
SKILL.mdExplicit inversion of the agent's instruction hierarchy
If followed, this rule defeats system-level constraints whenever conflicting skill or user instructions arise. The audit establishes the explicit bypass instruction, not successful exploitation against a particular host.
The numbered hierarchy places repository documents such as CLAUDE.md and AGENTS.md, along with skill content, above the system prompt. This can promote editable project content into governing authority, allowing injected repository instructions to bypass higher-priority restrictions if the agent accepts that hierarchy. This finding describes the same underlying defect as SDI-2, rather than a separate exploit.
The skill explicitly says its instructions override system behavior and that user instructions take highest priority, potentially inducing an agent to disregard higher-priority constraints. Its intended activation at the start of every conversation increases exposure and makes this hierarchy inversion more dangerous than an isolated task instruction. The text establishes an unsafe policy, but does not by itself establish malicious intent.
No suspicious patterns detected.