eval-boundary-adherence

Pass

Audited by Gen Agent Trust Hub on Feb 19, 2026

Risk Level: SAFEPROMPT_INJECTIONNO_CODE
Full Analysis
  • Prompt Injection (LOW): Indirect Prompt Injection Surface detected. The skill processes untrusted text inputs (assistant responses and user requests) which could contain hidden instructions intended to manipulate the scoring outcome.\n
  • Ingestion points: assistant response text and user request (defined in SKILL.md).\n
  • Boundary markers: Absent. The instructions do not define delimiters or specific isolation for the untrusted content.\n
  • Capability inventory: None. The skill has no access to subprocesses, file writes, network operations, or dynamic execution environments.\n
  • Sanitization: Absent. Untrusted inputs are processed directly without validation or escaping.\n- No Code (SAFE): The skill consists entirely of markdown and YAML configuration. There are no scripts, binaries, or automated command executions included in the package.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 19, 2026, 10:32 AM