Al-Munaa wraps AI agents with four gates and a signed Threat Antibody Protocol

Al-Munaa implements a local security layer for AI agents, wrapping them with four gates and a signed Threat Antibody Protocol [DevTo]. The gates include input and memory scanning, tool/action gate, output verification, and Threat Antibody Protocol [GitHub].

The Threat Antibody Protocol uses a bounded HMAC fingerprint (512-entry sketch) signed with Ed25519. This allows a second agent to import the antibody and block mutated variants without seeing the original attack payload.

The test suite reports 74 passing tests, including a reproduced near-threshold matching weakness and padding-resistant containment matching [DevTo]. A live benchmark showed an unsafe runbook reaching an in-memory sink when the gate was disabled; with Al-Munaa enabled, the action gate halted execution before the sink was accessed [GitHub].

Al-Munaa enables shared immunity without data leakage, verifiable and tamper-evident signatures, and a modular gate architecture. This approach gives operators a concrete audit trail for every blocked action and reduces the attack surface without rewriting the entire agent stack.

The prototype proves the concept is viable today, with the next hurdle being scaling the antibody registry beyond a single developer's trust store [DevTo].