colab-gpu
Pass
Audited by Gen Agent Trust Hub on Mar 12, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill documentation instructs the agent to use the Bash tool for executing local synchronization scripts (e.g.,
./scripts/colab_sync.sh) and training code (python train.py). - [PROMPT_INJECTION]: The skill implements an automated monitoring system that ingests an externally generated JSON file (
_colab_complete.json) from Google Drive/Colab, creating an indirect prompt injection surface. - Ingestion points: Reads
_colab_complete.jsonfrom the results directory synced from Google Drive. - Boundary markers: None identified in the provided instructions.
- Capability inventory: The skill has access to
Bash,Read, andGreptools, allowing for file execution and manipulation. - Sanitization: There is no evidence of validation or sanitization of the JSON content before it is processed.
Audit Metadata