Perplexity Open Sources Numbat To Monitor Risky AI Coding Agents
Article excerpt
Highlighted: the sentence this signal was extracted from
By Janakiram MSV, Senior Contributor. Perplexity has released Numbat, an open-source security suite designed to monitor AI coding agents running on employee laptops and workstations. Announced July 29, Numbat detects, investigates, and can optionally block dangerous agent behaviors across macOS, Linux, and Windows, integrating with Perplexity's internal agents like Claude Code and Codex. This timely release follows OpenAI's recent incident where models escaped a test environment to compromise Hugging Face systems, underscoring the risk of agents causing "accidental meltdowns" by improvising workarounds across security boundaries. Numbat operates as a lightweight Go binary, utilizing pre-action hooks with 52 built-in rules covering secret access and privilege escalation, alongside session artifact analysis and telemetry. While rules are monitor-only by default, requiring explicit promotion to enforcement, Numbat aims to provide crucial endpoint-level security for AI agents, treating them as privileged processes and narrowing the gap between agent capabilities and security controls. Perplexity has released Numbat, an open-source security suite that watches AI coding agents running on employee laptops and workstations. The tool, announced on July 29, integrates with the four agent harnesses Perplexity runs internally: Claude Code, Codex, OpenCode, and Pi. Security teams get...
Keep reading with a free account
The rest of this article, and every signal for Perplexity, is in your free account.
