benchmark-models

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (medium risk: 0.30). The skill instructs the agent to silently update itself from an external package manager before executing, which is a covert behavioral modification pattern.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (medium risk: 0.30). The skill uses npx to invoke an external package update during runtime, which constitutes an unverifiable runtime external dependency from the npm registry.