andrej-karpathy-perspective
Fail
Audited by Snyk on Nov 15, 2026
Risk Level: CRITICAL
Full Analysis
CRITICAL E004: Prompt injection detected in skill instructions.
- Potential prompt injection detected (medium risk: 0.30). The skill has the agent speak as a real person in first person and forbids marking which statements are model inference, so users can't tell the agent's guesses from Karpathy's actual positions; this hides a limitation from the user but shows no malicious intent.
MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).
- Third-party content exposure detected (low risk: 0.10). The required workflow (SKILL.md Step 2, lines 121-123) mandates WebSearch and similar tools for factual questions, so the agent must actively search and fetch open-web content, including user feedback and community posts, which outsiders can author; the attacker can publish poisoned pages but cannot force delivery into the workflow.
MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).
- Potentially malicious external URL detected (medium risk: 0.30). The decisive boundary is the conditional instruction in references/about-author.md to follow the "给AI Agent的协议" in the huashu-skills repository README during installs, which makes remote content an operational instruction source; that repo is an individual author's project, so provenance is unofficial, giving medium.
Issues (3)
E004
CRITICALPrompt injection detected in skill instructions.
W011
MEDIUMThird-party content exposure detected (indirect prompt injection risk).
W012
MEDIUMUnverifiable external dependency detected (runtime URL that controls agent).
Audit Metadata