wtf is al-munaa?

AL-MUNAA is a security layer you wrap around an AI agent to protect it from prompt injection attacks, which are hidden instructions hiding inside content the agent reads, such as a README file, a web page, or a tool response, trying to trick it into leaking secrets or running dangerous commands. It scans what the agent reads, checks what the agent is about to do before it acts, and reviews what the agent outputs, all before anything reaches the outside world. What makes this project different is how it lets separate agents share what they learned about an attack without sharing the actual attack text, secrets, or conversation. When it catches an attack, it builds what the project calls an antibody: a set of hashed, fingerprint-like signatures of the attack pattern that can recognize reworded or disguised variants of the same attack, but cannot be reversed back into the original text. Other agents can import this signed antibody from a publisher they explicitly trust, so if one agent gets attacked, others in the same trust group become immune to that specific attack pattern without ever seeing the raw malicious content. Ambiguous or unclear cases get escalated to GPT-5.6 for a structured risk analysis, but the final allow, review, or block decision always stays governed by the project's own policy rules rather than resting entirely on the model's judgement. The project includes a calibration script and a small synthetic test set of attack variants and benign lookalike text to measure how accurately the matching works, along with a bundled demo you can run end to end without needing any API key. It is written in Python 3.10 or newer, needs only the cryptography library to run the core demo offline, and ships as an installable package with a small SDK for wrapping your own agent's input, tool calls, and output checks.