Numbat Tool Enhances AI Agent Observability and Security
Perplexity AI's open-source tool, numbat, provides critical observability for AI agents, addressing concerns about uncontrolled proliferation and excessive privileges.
Enterprises are grappling with an unmanaged crisis of AI agent and MCP server sprawl, characterized by the rapid, decentralized proliferation of autonomous agents. These agents often operate with excessive privileges, possess opaque execution paths, and create identity blind spots. Without agent-aware governance, organizations struggle to prevent, detect, or contain multi-hop autonomous exploits, leaving environments vulnerable to lateral movement, shadow collaboration, and unauthorized data exfiltration. The recent OpenAI/Hugging Face incident highlights that monitoring capabilities are lagging behind agent advancements, with tooling to observe real-time agent activities not yet standard practice, even within AI development labs.
To address this gap, Perplexity AI has released numbat, an open-source tool designed to offer enhanced observability and visibility for AI agents. Numbat supports desktop, CLI, and IDE environments through local hooks and plugins, leveraging OTLP/HTTP logs and on-disk session artifacts. The tool is distributed as a Go binary, compatible with macOS, Linux, and Windows, and can be installed easily.
Numbat simplifies agent discovery with commands like ./numbat agents, which provides a detailed breakdown of installed agents, their configurations, available artifacts, and hook status. Users can install hooks for supported agents, such as Claude and Gemini, using simple commands like ./numbat hook install --agent claude --emit all. Once hooks are active, numbat can log various behaviors, writing events, findings, and indicators to a local file for analysis.
The tool includes a comprehensive catalog of detection rules categorized by behavior types like secrets, exfiltration, integrity, execution, reconnaissance, and privilege escalation. For instance, the recon.network_sweep detector, which identifies network scanning activities, can be triggered by a simple prompt. When activated, numbat logs findings to ~/.numbat/records.ndjson, allowing for detailed forensic analysis.
Beyond detection, numbat supports enforcement actions to prevent unauthorized activities. Administrators can configure rules to block specific behaviors by setting enforce: true in their YAML definitions. This allows for proactive security measures, preventing agents from performing actions deemed risky or unauthorized within the enterprise environment.
Numbat also facilitates incident response investigations by packaging relevant artifacts. Users can create case IDs, output investigation data, and build case packages that include event logs, findings, and a manifest file with SHA256 hashes for integrity verification. This feature streamlines the process of gathering evidence and understanding the scope of a security incident involving AI agents.
Furthermore, numbat offers a timeline view of agent activities, grouping events by agent, type, and session ID for chronological analysis. The tool also supports advanced deployment options, including live OTLP/HTTP capture and asynchronous HTTP delivery, with opportunities for managed configuration via MDM for fleet-wide deployments.
Numbat's capabilities are crucial for organizations seeking to manage the security risks associated with the growing adoption of AI agents. By providing visibility into agent actions, enabling enforcement, and supporting incident response, numbat helps close a critical gap in AI security governance.