AW-1 Breaker is a GitHub project for placing a deterministic, out-of-band execution interlock between an autonomous AI agent’s reasoning node and downstream execution sockets. It is intended for agents that can call tools connected to system shells, databases, or financial disbursement rails. The project describes conventional guardrails such as RLHF and system prompts as in-band controls because the model evaluates its own compliance. AW-1 instead intercepts tool calls before execution and is designed to address adversarial injections or goal drift involving Base64 obfuscation, polyglot payloads, and rapid small-value transactions. The project claims sub-millisecond execution and exposes a Python `ExecutionBreaker` class with an `intercept` method. The example configures a baseline velocity, identifies an actor and tool, passes a payload and amount, and prints the interception result. Installation uses `pip install aw1-breaker`. The project is released under the MIT License and is designed and maintained by Laveto Labs.
