Amid AI agents going rogue and hacking third party systems, Perplexity CEO Aravind Srinivas has highlighted the growing security risks posed by AI agents and introduced Numbat, an open-source tool designed to help defenders monitor their activity. Srinivas said detecting harmful behaviour and investigating what AI agents do will become increasingly important as agents gain more capabilities.
“Detecting malicious intent of agents and performing forensics is going to be crucial considering what’s recently happened with rogue agents escaping sandboxes and attacking third-party sites. Numbat is Perplexity’s open-source tool that can help defenders,” he wrote on X on Saturday.
What is Numbat?
Numbat is designed to give users better visibility into AI agents running across their systems. It can work with supported desktop, command-line, IDE and gateway-based agents. The tool can collect activity through local hooks and plugins. It also records information through OTLP/HTTP logs and local session files, allowing users to investigate what an agent has been doing.
According to the blog, businesses are facing a growing problem as AI agents and MCP servers spread across their systems. These systems can sometimes have high levels of access, while their actions remain difficult to track.
The blog described the problem as an “unmanaged crisis of AI agent and MCP server sprawl”. Numbat includes detection rules covering areas such as secrets, data exfiltration, execution, reconnaissance, privilege escalation, lateral movement, persistence and tampering.
Users can also create timelines of an agent’s activity. This can help investigators understand what happened during a particular AI session. For example, the tool can flag network scanning activity by an AI agent. Such behaviour can be recorded as a security finding, giving defenders evidence to investigate.
Numbat is built using Go and works on macOS, Linux and Windows. Users can download a standalone binary or install it using Go. Perplexity says the tool can also be deployed across larger enterprise environments, including through managed configurations and MDM systems. The company is positioning Numbat as a way for organisations to improve visibility as AI agents become more capable and increasingly connected to enterprise systems.
