OpenLeash adds human approval for risky AI agent actions

OpenLeash is a monitoring and authorization layer for autonomous AI agents developed by Max Brin. It inspects agents’ intended actions, pauses or blocks operations judged risky, and requests human approval when actions are uncertain.

Brin developed the tool because agents often inherit a user’s permissions without inheriting human situational awareness. That combination can let a single bad prompt, a compromised model or a malicious tool cause damage across a network.

OpenLeash runs alongside in-house agents, cloud agents and third-party agents. It captures an agent’s intended request before it executes, applies configured policy checks, and either allows, blocks or pauses the action pending operator input.

Administrators can configure acceptable API endpoints, allowed destinations and payment limits. The system can automatically approve payments below a set threshold while requiring human authorization for larger transfers. Configuration settings can be changed at any time.

The product is under active development and already in use. Brin reports several hundred individual users and at least four organizations running OpenLeash. He expects a set of planned feature updates to arrive in the coming months.

Brin describes the product as “AV for AI” and calls it “medication for AI anxiety.” He offers an example: when an agent attempts to delete a database, OpenLeash intercepts the request, evaluates risk and either blocks the action or asks the operator whether to proceed.

Security teams have sought tools that add permission checks as agents take more real-world actions. OpenLeash provides a policy-driven permission layer designed to reduce unapproved operations such as destructive deletes or credential exfiltration.