Last month, an agentic AI assistant called OpenClaw that promised to manage your calendar, check you in for flights, respond to emails, and organize your files went viral. Within weeks, security researchers had found over 30,000 exposed instances on the internet. A Meta AI safety researcher had watched helplessly as it deleted her inbox before she could stop it, and the tool became the example of what ungoverned AI looks like in practice.
For the founders of JetStream Security, a San Francisco–based startup built by veterans of CrowdStrike, SentinelOne, and Cohesity, it’s an example of the problem they’re trying to solve. Companies are racing to deploy AI agents and custom-built models, but most have no way to map what those systems are doing, no inventory of unauthorized AI tools their employees are quietly running, and no kill switch for when something goes wrong.
