evals-implement

Warn

Audited by Gen Agent Trust Hub on Aug 4, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill dynamically generates Python grader scripts and unit tests in the evals/ directory based on contents from goldset.json. It implements a self-tuning mechanism in Phase 2b that edits and re-executes these scripts up to three times.\n- [COMMAND_EXECUTION]: The skill executes the pytest command to run the generated unit tests against the generated grader logic.\n- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting external data to synthesize LLM-judge templates and code logic.\n
  • Ingestion points: Reads criteria and examples from evals/{system}/goldset.json in Phase 1.\n
  • Boundary markers: No delimiters or instructions are used to prevent the agent from obeying instructions embedded within the goldset data.\n
  • Capability inventory: Subprocess execution of pytest and file system write access in evals/{system}/graders/ and evals/{system}/tests/.\n
  • Sanitization: No validation or sanitization is applied to the goldset data before it is interpolated into executable code or prompts.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 4, 2026, 03:33 PM
Security Audit — agent-trust-hub — evals-implement