Show HN: Open Bias – proxy that enforces agent behavior at runtime (github.com)

Open Bias has been released as an open-source agent alignment tool that acts as a proxy between applications and large language model (LLM) providers, offering real-time enforcement of user-defined rules without latency or complex configuration. Users can simply install Open Bias and point their application to its proxy, where it intervenes to correct off-policy behavior before it reaches the end users. The tool leverages a Markdown file for rules, making it accessible for teams to collaborate on policy management, and includes multiple evaluators that can address various enforcement needs. The significance of Open Bias for the AI/ML community lies in its ability to address common shortcomings with LLMs, such as the unreliability of models following complex prompts and rules. In real-world applications, models often exhibit undesired behaviors, such as revealing sensitive information or failing to execute safety protocols. Open Bias evaluates requests at runtime, blocking or modifying responses that violate established rules, which enhances the safety and compliance of AI interactions. This tool simplifies the enforcement of policies, helping developers maintain robust control over agent behavior while ensuring smooth operations with zero added latency.