CrabTrap: An LLM-as-a-judge HTTP proxy to secure agents in production

Article

• HTTP proxy from Brex that intercepts agent traffic for security review
• Uses an LLM as judge to evaluate requests/responses before they go through
• Designed to catch prompt injection and malicious agent behavior in prod

Discussion

• One commenter is building a competing approach in the same space
• Skeptic argues using LLMs to secure LLMs adds risk, not reduces it
• Consensus: agent security is underserved, but LLM-judge approach is debated

Discuss on HN