The original title is "AL-MUNAA: a collective immune system for AI agents"

Short summary

AL-MUNAA is an open-source local security layer for AI agents featuring input scanning, tool/action gating, output verification, and a signed Threat Antibody Protocol for cross-agent immunity. When one agent detects a prompt injection, it generates an Ed25519-signed HMAC fingerprint that other agents can use to block mutated variants without sharing raw attack text. The project includes 74 passing tests and was built for OpenAI Build Week, with a demo showing it blocking an unsafe runbook from reaching a synthetic data sink.

• •Four-gate security layer: input scanning, action gating, output verification, and signed threat antibody protocol
• •Agents share Ed25519-signed HMAC fingerprints to immunize each other without exposing private data
• •74 passing tests; demonstrated blocking an unsafe runbook from reaching a synthetic in-memory sink

Generated with AI, which can make mistakes.