grpo-rlvr-training

Warn

Audited by Gen Agent Trust Hub on Jul 14, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The test_execution_reward function in references/reward-functions.md executes model-generated completions as Python scripts. This creates a surface where model-generated content can execute logic on the host or in a sandbox.
  • [COMMAND_EXECUTION]: The skill uses subprocess.run to invoke the pytest runner. It implements several security controls as part of its execution flow: [1] Ingestion points: Untrusted model-generated completions are processed in references/reward-functions.md. [2] Boundary markers: The skill instructions in SKILL.md and regex patterns in references/reward-functions.md encourage the use of reasoning and answer tags to isolate executable content. [3] Capability inventory: The skill performs file writes to temporary directories and executes subprocesses in references/reward-functions.md. [4] Sanitization: The function performs environment scrubbing by restricting the PATH and requires a mandatory sandbox_cmd prefix for execution, refusing to run if an isolation boundary is not provided.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 14, 2026, 07:33 PM
Security Audit — agent-trust-hub — grpo-rlvr-training