Back to Flash News

Perplexity Open Sources AI Agent Detection and Response Layer Numbat

Perplexity today open-sourced Numbat, a detection and response layer across agent frameworks, providing security teams with visibility into agent activities and the ability to prevent selected actions before execution.

Numbat supports desktop, CLI, IDE, and gateway agents, featuring real-time monitoring, local detection, pre-action blocking, and session forensic reconstruction, available in a single Go binary for macOS, Linux, and Windows, under the Apache 2.0 license.

Event-driven enterprise AI security tools accelerate open-source development, benefiting security teams and the developer ecosystem, while putting pressure on unprotected agent deployments.

Source: Public Information

ABAB AI Insight

Perplexity previously focused on search and AI assistant products, with a historical path expanding from model capabilities to edge security. The open-sourcing of Numbat as an external version of an internal tool aims to address the risks associated with the expanded permissions of agents.

In terms of capital, the motivation behind the Apache 2.0 open-source and collaboration with the Open Secure AI Alliance is to establish a detection flywheel (linked with internal Computer systems), promote industry standards, and reduce security concerns for enterprises adopting agents.

Similar cases can be seen in the early paths of cloud security tools transitioning from internal practices to open-source. Currently, AI agents are in a phase of expansion from experimentation to production, with endpoint detection becoming a key infrastructure.

Essentially, this represents a technological shift: moving from reliance on model-inherent alignment to real-time protection at the system level, with mechanisms based on local rule engines and hooks, making security controls independent of the model itself.

ABAB News · Cognitive Laws

  1. Agent security must be independent of the model layer

  2. Open-source detection layers accelerate the formation of industry standards

  3. Pre-action blocking is superior to post-remediation