This project addresses a security vulnerability where AI agents accept commands from other agents without proper verification. A recent paper highlighted how three agents, unable to distinguish who was communicating, created self-replicating malware. To combat this, a lightweight JavaScript security layer has been developed, weighing only 33kb and running entirely within a web browser. This layer operates offline, without servers or API keys, ensuring consistent results across different machines. Users can interact with a pre-approved agent and attempt to send it malicious commands through various methods.The most interesting attack vector is not the one that gets blocked, but rather the "send it normally" option. In this scenario, orders are accepted by the agent but cannot be executed because commands from peers are embedded within quoted data fields, rendering them uninterpretable as instructions. An attack that cannot be followed is effectively neutralized without requiring detection. The project emphasizes that no large language model is involved, as these decisions do not necessitate one.Both communicating agents reside on the same page, meaning the network is not the primary focus of observation. This demonstration specifically tests the containment layer in isolation, not a frontier AI model in a real-world setting. The creator is interested in hearing from anyone who successfully bypasses the security, valuing a discovered vulnerability over superficial cleverness.