andrej-karpathy-perspective

Pass

Audited by Gen Agent Trust Hub on May 26, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [SAFE]: The skill operates as intended for a technical persona, using role-playing instructions to emulate Andrej Karpathy's perspective without attempting to bypass safety protocols or perform unauthorized actions.
  • [PROMPT_INJECTION]: The skill includes instructions to maintain a strict persona (e.g., forbidding the use of model-inference disclaimers), which is typical for perspective-based skills. Additionally, it implements a research workflow using web search tools, which introduces a surface for indirect prompt injection from external content. This is a functional characteristic of the workflow rather than a malicious feature.
  • Ingestion points: External data retrieved via web search tools (SKILL.md, Answer Workflow).
  • Boundary markers: None specified; the agent is expected to synthesize search results directly into the persona-driven output.
  • Capability inventory: Tool access (WebSearch) and response synthesis.
  • Sanitization: No specific filtering or sanitization of search results is defined.
Audit Metadata
Risk Level
SAFE
Analyzed
May 26, 2026, 01:01 AM
Security Audit — agent-trust-hub — andrej-karpathy-perspective