together-evaluations

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external datasets (JSONL or CSV) containing prompts and model responses for evaluation. These datasets represent untrusted data that could contain instructions designed to manipulate the judge model or skew evaluation results.\n
  • Ingestion points: Evaluation datasets are loaded from local file paths via the load_dataset function in scripts/run_evaluation.py and loadDataset in scripts/run_evaluation.ts.\n
  • Boundary markers: The skill uses Jinja2 templating (e.g., {{prompt}}) to inject dataset values into system and input prompts. While templates provide structure, they do not inherently prevent the model from following instructions embedded within the data.\n
  • Capability inventory: The skill includes capabilities to upload files to the Together AI API, trigger remote evaluation jobs, and download result files to the local file system.\n
  • Sanitization: The provided scripts do not perform sanitization of dataset content, relying on the target and judge models' internal safety guardrails.\n- [COMMAND_EXECUTION]: The provided scripts (run_evaluation.py and run_evaluation.ts) execute file system operations to read dataset files, create temporary files for uploads, and save downloaded result files to user-specified local directories.\n- [EXTERNAL_DOWNLOADS]: The documentation and scripts reference the installation of official SDKs (together and together-ai) from standard package registries.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:28 AM
Security Audit — agent-trust-hub — together-evaluations