nemo-gym-blade-analysis
Pass
Audited by Gen Agent Trust Hub on Jul 22, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill processes untrusted rollout trajectories and logs, which creates a surface for indirect prompt injection attacks.
- Ingestion points:
scripts/blade_toolkit.py(processes rollout JSONL records and report markdown files). - Boundary markers: The skill relies on natural language instructions for the agent to "read trajectories in order" but does not implement automated delimiters or safety boundaries within the toolkit scripts.
- Capability inventory: The
blade_toolkit.pyscript performs file system operations (read/write) and data extraction; the agent environment allows shell command execution. - Sanitization: The skill provides manual sanitization guidelines in
references/cvdp-report-example.mdto remove credentials and private data before report generation. - [COMMAND_EXECUTION]: The skill uses
uv run pythonto execute the localscripts/blade_toolkit.pyscript for benchmark validation, fact extraction, and scoring. The script is bundled with the skill and uses standard libraries to process local files.
Audit Metadata