Perplexity has open-sourced Numbat, a lightweight security layer for AI coding agents running on developer machines. Released under the Apache 2.0 license around 29-30 July 2026, it was contributed to the Open Secure AI Alliance, which includes groups such as NVIDIA. Numbat ships as a single static Go binary that runs on macOS, Linux and Windows for both amd64 and arm64. It works as a cross-harness tool, pulling activity from saved sessions, live hooks and log streams into one common event format. A local rule engine based on Common Expression Language can evaluate actions in real time. When hooked into an agent's pre-action system, it can block high-risk steps such as destructive shell commands or data leaks before they run. Blocking is off by default and must be turned on. For later review, Numbat rebuilds full session timelines from files already on disk and packages them into portable case bundles with secret redaction. Everything stays on the local machine.
Perplexity has open-sourced Numbat, a lightweight security layer for AI coding agents running on developer machines. Released under the Apache 2.0 license around 29-30 July 2026, it was contributed to the Open Secure AI Alliance, which includes groups such as NVIDIA. Numbat ships as a single static Go binary that runs on macOS, Linux and Windows for both amd64 and arm64. It works as a cross-harness tool, pulling activity from saved sessions, live hooks and log streams into one common event format. A local rule engine based on Common Expression Language can evaluate actions in real time. When hooked into an agent's pre-action system, it can block high-risk steps such as destructive shell commands or data leaks before they run. Blocking is off by default and must be turned on. For later review, Numbat rebuilds full session timelines from files already on disk and packages them into portable case bundles with secret redaction. Everything stays on the local machine.