evidence-draft

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFE
Full Analysis
  • [Prompt Injection] (SAFE): The instructions provide clear role-based guidance for data processing. No bypass markers or 'ignore previous instructions' patterns were found.
  • [Data Exposure & Exfiltration] (SAFE): The skill is restricted to project-specific paths (papers/, outline/, citations/). The description explicitly states no network access is used. No hardcoded credentials or sensitive system paths were detected.
  • [Unverifiable Dependencies & Remote Code Execution] (SAFE): The skill references a local Python script (.codex/skills/evidence-draft/scripts/run.py) for processing. No remote scripts (e.g., via curl/wget) are downloaded or executed.
  • [Obfuscation] (SAFE): No encoded strings (Base64), zero-width characters, or homoglyphs were identified in the skill definition.
  • [Indirect Prompt Injection] (SAFE): While the skill ingests external data (paper_notes.jsonl), the strict 'NO PROSE' requirement and structured JSONL output format serve as inherent mitigations against instructions embedded in research content.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 06:06 PM