I build an open-source runtime security layer for AI agents – Let's break it (github.com)

A new open-source project, Pryxor, has been launched to enhance security for AI agents by acting as a runtime security layer. AI agents, even with valid credentials, can perform unauthorized or unwanted actions, such as accessing sensitive data or executing unnecessary commands. Pryxor addresses this by evaluating each action proposed by an agent before it interacts with external systems, deciding whether to approve, block, or hold actions for human review. This explicit and deterministic control allows organizations to establish clearer security policies tailored to specific actions. Pryxor is notable for its capability to isolate agent credentials and ensure actions are logged and auditable, enhancing overall security postures for AI deployments. It enforces policies through a structured approval lifecycle, requiring human intervention for sensitive actions, and is designed for scenarios where agents can perform various tasks, including sending emails and accessing APIs. While still in early development, Pryxor provides fundamental tools for defining and testing custom security policies, inviting contributions from the community to enhance its functionalities, such as robust action validation and policy simulations.