evaluating-code-models

Warn

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: Fetches the evaluation harness from the BigCode Project's official GitHub repository and downloads Docker images from the GitHub Container Registry (ghcr.io).
  • [COMMAND_EXECUTION]: Instructs users to enable the --allow_code_execution flag, which permits the execution of arbitrary, untrusted code generated by AI models during the benchmarking process.
  • [DYNAMIC_EXECUTION]: Recommends the use of the --trust_remote_code flag when loading specific models from HuggingFace, which allows the execution of arbitrary Python code provided by the model's repository.
  • [PRIVILEGE_ESCALATION]: Includes instructions for building Docker containers using sudo in the troubleshooting documentation (sudo make DOCKERFILE=Dockerfile all).
  • [INDIRECT_PROMPT_INJECTION]: The skill operates as a pipeline for processing and executing external model outputs, creating an attack surface for malicious code injection.
  • Ingestion points: Model-generated code is loaded from JSON files (generations.json) or directly from model inferences for evaluation.
  • Boundary markers: No explicit boundary markers are used to separate instructions from the code being executed; the system relies on user-provided flags for execution control.
  • Capability inventory: The evaluation script (main.py) has the capability to execute shell commands and arbitrary code through subprocesses and the Python interpreter.
  • Sanitization: The documentation provides mitigation strategies, specifically recommending the use of Docker containers to isolate the host system from the code execution environment.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 17, 2026, 07:53 PM
Security Audit — agent-trust-hub — evaluating-code-models