poles
Warn
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill instructs the agent to dynamically load and prioritize instructions from files located at paths defined by the environment variable '$POWERSUIT_OVERLAY', specifically seeking a 'README.md' and files within a 'stance/poles/' directory.- [COMMAND_EXECUTION]: The instructions provide a specific shell command block '[ -n "$POWERSUIT_OVERLAY" ] && ls -1 "$POWERSUIT_OVERLAY/stance" 2>/dev/null || true' for the agent to execute to verify the existence of local configuration overlays.- [INDIRECT_PROMPT_INJECTION]: The overlay mechanism creates a vulnerability where instructions can be injected into the agent's context from unverified local files. Ingestion points include the '$POWERSUIT_OVERLAY/stance/' directory. No boundary markers or sanitization procedures are defined for this external content, and the skill explicitly directs the agent to follow the overlay's instructions over its own where they disagree.- [PROMPT_INJECTION]: The skill directs the agent to 'ignore base rates entirely' and act in a 'delusional' manner when constructing specific parts of its response, which is a directive to bypass standard reasoning and logical constraints.
Audit Metadata