training-check

Warn

Audited by Gen Agent Trust Hub on Oct 23, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill is explicitly instructed to execute a stop_command retrieved from the session context and is directed to manage processes by terminating tmux sessions, local processes, remote processes via SSH, and scheduler jobs.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources, including WandB metrics and training log files. Content within these logs could potentially be manipulated to influence the agent's decision logic, such as triggering an unauthorized training shutdown.
  • Ingestion points: WandB run URLs and local or remote training logs as described in SKILL.md.
  • Boundary markers: The instructions lack specific delimiters or warnings to prevent the agent from obeying instructions embedded in the logs.
  • Capability inventory: The skill utilizes the Bash tool and has instructions to terminate processes and write files.
  • Sanitization: There are no instructions for validating or filtering metric data or log content before processing.
  • [DYNAMIC_EXECUTION]: The skill relies on an external script reviewer_route.py and a local configuration file .aris/reviewer.json for managing its own invocation routing and deterministic gating, creating a dependency on unverified execution logic.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 23, 2026, 06:50 AM