Strengthening Agent Security for Perplexity Client Endpoints with Numbat

As AI agents gain greater autonomy, security incidents arising from agents performing actions that diverge from user intent—even without external attacks (Prompt Injection)—are emerging as a new threat.

Since this problem cannot be solved simply by patching the model itself, defense at the Agent Harness stage—the interface connecting the model to the outside world—is essential.

Numbat is an open-source security suite designed to detect, investigate, and prevent risky activity from agents running on enterprise client endpoints. Its key features are as follows:

• Unified Interface: Integrates directly with a variety of client-side agent harnesses, reducing the burden on security teams of having to build individual protection tools for each agent.
• Local-First Detection: Performs detection directly at the endpoint and provides forensic capabilities that are agnostic to the type of agent.
• Risk Management: Monitors and controls in real time the permission abuse and autonomous misbehavior that can arise from settings such as --dangerously-skip-permissions.

Numbat was released as part of the Open Secure AI Alliance, alongside NVIDIA and others, and helps safely manage the elevated privileges that agents hold in enterprise environments.