spatial-reasoning
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documentation in advanced.md describes a mandatory protocol involving a persistent command loop that polls /tmp/gtc2_cmd.json for instructions. This creates a potential surface where an attacker with write access to this temporary file could inject malicious commands that the agent would then process. Ingestion points: /tmp/gtc2_cmd.json described in advanced.md. Boundary markers: None identified in the provided content. Capability inventory: run_script execution helper and described 'exec' actions for stage edits. Sanitization: No sanitization or validation logic for the JSON command content is described.
- [COMMAND_EXECUTION]: The documented 'HARD RULE' for executing stage edits through an 'action: exec' payload within a command file creates a significant risk of arbitrary code or command execution if the runtime implementation does not strictly validate the input source and content.
Audit Metadata