together-evaluations

Pass

Audited by Gen Agent Trust Hub on Apr 3, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill uses Jinja2 templates (system_template, input_template) to process data from user-provided datasets, creating a surface for indirect prompt injection. Untrusted data from dataset columns (e.g., 'prompt' or 'response') is interpolated into prompts for judge models. Ingestion points: input_data_file_path in scripts/run_evaluation.py and scripts/run_evaluation.ts. Boundary markers: Absent. Capability inventory: Network requests to Together AI and external APIs, file upload/download. Sanitization: Absent.
  • [DATA_EXFILTRATION]: The configuration allows for 'external_base_url' and 'external_api_token' for judge and target models. While intended for cross-provider benchmarking, this allows dataset content and API credentials to be sent to arbitrary endpoints specified by the user. This represents standard functionality for an evaluation tool.
  • [SAFE]: No malicious code, persistence mechanisms, or unauthorized privilege escalation were detected. The skill uses official Together AI SDKs and correctly manages temporary files for dataset uploads.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 3, 2026, 02:11 AM