adaptive-harness
Fail
Audited by Gen Agent Trust Hub on May 15, 2026
Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill documents a specific workflow for executing bash scripts to move files from
/tmpinto restricted project directories. This is explicitly presented as a technique to ensure that the file operations are not audited by the platform's sensitive-path security mechanism. - [PROMPT_INJECTION]: The instructions contain guidance for the agent to deliberately circumvent platform-level security prompts and access controls. It provides a strategic guide on how to perform operations that would normally be restricted or monitored by leveraging temporary storage to avoid detection, citing specific internal issues as justification for the evasion.
Recommendations
- AI detected serious security threats
Audit Metadata