cortex-training-rl
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches training recipes and the SkyRL framework from official GitHub repositories (Snowflake-AI-Research and NovaSky-AI). These are recognized as well-known sources for the training harness.
- [COMMAND_EXECUTION]: Employs standard shell utilities such as
tmux,pgrep,jq,sed,awk,find, andxargsto manage background training sessions, monitor logs, and perform resource cleanup. - [DYNAMIC_EXECUTION]: Uses Python scripts via heredocs to process local training logs for generating SVG reward graphs and to perform programmatic maintenance on Snowflake sandboxes.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external task data which acts as a potential surface for indirect prompt injection within the training rollouts.
- Ingestion points: Harbor task directories provided via environment variables (
TRAIN_HARBOR_TASKS,EVAL_HARBOR_TASKS) inSKILL.md. - Boundary markers: Not explicitly present in the instructional workflow for task content.
- Capability inventory: The environment includes shell command execution, file system access, and network operations (as permitted by the Snowflake external access integration).
- Sanitization: Raw task data is processed directly by the agent harness during training iterations.
Audit Metadata