clj-skill-eval
Fail
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: HIGHPRIVILEGE_ESCALATIONPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill explicitly instructs the user to launch subagents using the
--dangerously-skip-permissionsflag. This flag is designed to bypass the platform's security controls, allowing subagents to perform sensitive file operations, network requests, and tool executions without user approval or safety constraints. - [PROMPT_INJECTION]: The skill documentation includes sections such as 'Critical Permission Setting' which directs the user or agent to intentionally disable safety filters and bypass content guidelines. This overrides the platform's standard security posture.
- [INDIRECT_PROMPT_INJECTION]: The skill framework facilitates a workflow where subagents ingest untrusted data while operating with elevated privileges.
- Ingestion points: Subagents are instructed to read and execute instructions from a
TASK.mdfile (Phase 3). - Boundary markers: The skill lacks any delimiters or instructions for the subagent to ignore potentially malicious content within the task file.
- Capability inventory: Subagents are granted full tool access (including network and file system) via the recommended
--dangerously-skip-permissionsflag. - Sanitization: There is no validation or sanitization of the input
TASK.mdcontent before it is processed by the subagent.
Recommendations
- AI detected serious security threats
Audit Metadata