Details
- Perplexity announces Numbat, an open-source agent-detection and response layer designed to work across diverse AI agent harnesses.
- The system gives security teams visibility into agent activity and provides controls to block selected actions before they are executed.
- Numbat supports deployment across desktop, command-line (CLI), integrated development environment (IDE), and gateway agents, aiming for broad coverage of developer and production workflows.
- Core capabilities include live monitoring of agent behavior, local detection of suspicious activity, pre-action blocking, and forensic reconstruction using existing session artifacts.
- Numbat runs across Perplexity's systems and streams audit events, findings, and security alerts into Computer, the company’s telemetry analysis and security orchestration layer.
- Computer continuously analyzes Numbat telemetry, escalates suspicious behavior, and suggests new on-device detections, creating a feedback loop that improves Numbat’s detection quality over time.
- The software ships as a single Go binary, simplifying installation and cross-platform support for macOS, Linux, and Windows environments.
- Numbat is released under the Apache 2.0 license, enabling commercial use, modification, and redistribution within the open-source ecosystem.
Impact
By open-sourcing Numbat as a cross-platform Go binary with deep integration into its Computer telemetry system, Perplexity contributes to a growing focus on securing AI agents at the harness and workflow layer. This move aligns with broader industry efforts to add real-time monitoring and policy enforcement to agent frameworks, and may pressure rival AI tooling providers to offer similarly transparent, auditable control layers for agent operations.
