claude-api

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). The skill’s Agent SDK “WebFetch” path can fetch and analyze web pages at runtime based on user prompts (e.g., opening research URLs), meaning outsider-authored free text from fetched pages could be ingested without pre-selecting a specific trusted item.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly instructs using WebFetch to load live documentation (e.g., https://platform.claude.com/docs/en/about-claude/models/overview.md and https://raw.githubusercontent.com/anthropics/claude-agent-sdk-python/main/README.md) which will be fetched at runtime and can be injected into the model context (controlling prompts), and the agent examples also show executing remote packages via npx (e.g., @playwright/mcp@latest), so these remote URLs are runtime dependencies that can directly affect agent behavior or cause remote code to be run.