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.
JetStream’s answer is built around a feature called AI Blueprints—real-time graphs that map everything an AI system is doing inside an organization at any given moment. Each Blueprint traces the full chain of activity: which agents are running, which models they’re using, what data and tools they are interacting with, and who or what is behind each action. Rather than a static snapshot, Blueprints track live behavior, so if an AI system starts acting outside its intended purpose, the platform flags it. They also track cost, showing what each AI workflow is spending and who is responsible for it.
