gym-to-atif
Pass
Audited by Gen Agent Trust Hub on Oct 24, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted trace data from Gym rollouts to convert them into ATIF trajectories. It implements multiple defensive layers to mitigate potential injection risks: strict input size limits (128MB for sources, 25MB for ATIF), structural validation against the ATIF schema including recursion depth limits (max 8), and the inclusion of detailed provenance and uncertainty metadata in the output to inform downstream consumers.
- [EXTERNAL_DOWNLOADS]: The skill documentation refers to the 'nemo-compass' package for trace loading. This is a vendor-owned resource from nvidia-nemo and the skill instructions mandate manual installation of a specific pinned version (0.1.0), preventing automated external fetches during runtime.
- [COMMAND_EXECUTION]: The skill uses Bash to execute its internal Python scripts for data conversion. These scripts are designed for local, offline operation and include security features such as permission hardening (setting umask to 077 and using mode 0700/0600) for output directories and files, and the suppression of potentially sensitive trace content in CLI error messages.
Audit Metadata