msverl-daily-regression-triage
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes git commands using subprocess.run in scripts/sync_repos.py and scripts/list_recent_commits.py. These commands are used to clone, update, and inspect the commit history of the verl and MindSpeed repositories. The execution uses list-based arguments, which mitigates standard shell injection vulnerabilities.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from regression logs and git commit history, presenting this information to the agent for triage. An adversary with the ability to influence these logs or commit messages could attempt to inject instructions to influence the agent's behavior.
- Ingestion points: Regression logs at /home/st_daily_verl/msverl.log, training logs in /home/st_daily_verl/logs/, and commit metadata fetched via git log.
- Boundary markers: The skill does not employ explicit delimiters or system instructions to distinguish between log/commit data and operative agent commands.
- Capability inventory: The skill possesses capabilities for local file reading and executing git commands through subprocess.
- Sanitization: The skill uses ast.literal_eval for safe parsing of literal structures in logs and tokenizes commit messages for scoring purposes.
Audit Metadata