Formally verified protection against prompt injection and data exfiltration for AI agents, backed by Lean 4 proofs. The injection detector analyzes untrusted input and returns a verdict with probability score, bits at risk (via Adversarial Landauer bounds), and recommended action. The canon scanner checks code against a vulnerability database. Both run as hosted API endpoints with a generous free tier (1,000 detections monthly), though you can also grab the open source SDK and integrate via a simple POST. Reach for this when you're building agents that handle user input or execute code and need mathematical guarantees rather than heuristics. The Maxwell service adds adaptive proof of work defenses to make spam expensive for attackers.
claude mcp add --transport http io.github.viridis-security-injection-detector https://mcp.viridis-security.com/mcp