cloudflare-agents
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: Several templates, such as
BrowserAgentandRAGAgent, ingest untrusted data from external sources like scraped HTML and vector database context to be processed by an LLM. - Ingestion points: Scraped HTML content in
BrowserAgentand retrieved context inRAGAgent(e.g.,templates/browser-agent.tsandtemplates/rag-agent.ts). - Boundary markers: Prompt templates currently use simple string interpolation without explicit delimiters (e.g., XML tags) or instructions to ignore embedded commands.
- Capability inventory: The agents possess significant capabilities including network access to LLM providers, browser automation via Puppeteer, and persistent state management using SQLite.
- Sanitization: Input data is truncated for length but does not undergo sanitization to filter out potential malicious instructions embedded in the source content.
Audit Metadata