Perplexity releases Numbat open-source tool for AI agent security

Why Numbat was created

According to YourStory, Perplexity built Numbat to address growing concerns about AI agents operating with minimal supervision. AI agents—software systems that execute tasks beyond answering queries—can write code, run commands, access files and interact with external services. While this autonomy increases productivity, it also introduces security risks, particularly if agents access sensitive systems or operate for extended periods without oversight.

Perplexity’s CEO, Aravind Srinivas, highlighted the importance of detecting malicious intent and conducting forensic checks as agents become more capable. The tool aims to prevent incidents where agents escape sandboxes or attack third-party systems.

How Numbat monitors AI agents

Numbat tracks agent activity across desktops, command-line tools, integrated development environments (IDEs) and gateways. It collects data through local hooks, plugins, OTLP/HTTP telemetry logs and on-disk session files. This visibility allows security teams to reconstruct an agent’s actions during a session, according to YourStory.

The tool adopts a local-first approach, meaning activity logs remain on the device by default. Administrators can choose to process logs locally or forward structured telemetry to central systems for deeper analysis.

Security rules and forensic capabilities

Numbat includes 52 preconfigured rules across 11 behaviour categories, such as secret access, data exfiltration, privilege escalation, reconnaissance and tampering. For example, it can flag an agent attempting to access secrets and then transmit data outside the system as suspicious, as reported by YourStory.

The tool also supports forensic timelines, which help security teams understand the sequence of actions during an incident. This is critical because misconfigured or overly autonomous agents can create security issues without external attackers being involved.

Deployment and compatibility

Numbat is written in Go and supports macOS, Linux and Windows. It can run as a standalone binary or be integrated via Go, and enterprises can deploy it through managed configurations or mobile device management (MDM) systems, according to the source.

Perplexity uses Numbat internally to monitor its own coding agents, including Claude Code, Codex, OpenCode and Pi. The company is also part of the Open Secure AI Alliance, alongside NVIDIA and other organisations.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

Numbat fills a gap for startups and enterprises experimenting with AI agents but lacking the resources to build custom security monitoring. The tool’s open-source nature means teams can adapt it to their workflows without relying on proprietary solutions, though it does not replace broader AI governance policies. For founders, Numbat could be particularly useful if their products integrate autonomous agents or if they use agents internally for development or operations—areas where oversight is often sacrificed for speed.

Key takeaways

• Founders and operators using AI agents for automation or coding tasks now have a free, open-source tool to monitor and secure agent activity.
• Numbat’s local-first approach allows enterprises to retain activity logs on devices by default, reducing compliance risks.
• The tool’s built-in rules and forensic timelines can help teams detect misconfigurations or unintended actions before they escalate into security incidents.
• For startups testing autonomous agents, Numbat provides a way to balance productivity gains with oversight, especially if users bypass approval prompts for speed.