Cerbère – taint tracking for AI agents (github.com)

Cerbère, a new tool for AI agents, has been announced, focusing on runtime security and observability. This innovative solution aims to track the 'taint'—sensitive information that an agent might read and potentially misuse—by enforcing policy checks at the execution boundary. Unlike traditional prompt filtering, which only assesses what an agent is instructed to do, Cerbère scrutinizes every model and tool call to determine if it should be allowed, require human approval, or be blocked entirely. This capability is crucial for preventing actions like unauthorized access to sensitive data or unintended consequences of seemingly harmless commands. The significance of Cerbère lies in its multi-layered defense system that doesn't just stop at analyzing prompts but actively monitors the flow of data and actions. It includes features like advanced PII detection, tool usage policies, and human approval mechanisms, ensuring agents operate safely within predefined parameters. By utilizing machine learning for enhanced security measures, such as risk scoring and taint levels, Cerbère provides a robust framework for developers to maintain control over AI systems, mitigating risks associated with prompt injections, data exfiltration, and tool misuse while keeping human oversight in the loop.