Ken Kesey wrote about a ward where the patients ran things better than the staff did. Swap the ward for your AI stack and a lot of enterprises are operating on roughly the same model right now — minus Jack Nicholson's eyebrows, plus a much bigger cloud bill.

Could your company list every AI tool, model, and agent running across the business right now? Not the sanctioned ones on the vendor spreadsheet. The actual census — including whatever's still quietly running because the person who built it left in March and nobody wanted to touch it.

A few moments that stood out from a recent conference discussion:

  • One enterprise expected to find 70 AI tools and discovered 220. That's not a shadow IT problem anymore, that's group therapy — everybody's got a secret, and nobody's sharing in session.
  • Another uncovered an agent built by a former employee, still running at $1,800 a month, with nobody left who knew what it did. Not quite Frankenstein's monster — no villagers, no torches — just a recurring line item nobody's brave enough to cancel in case it's load-bearing.
  • From a legal perspective: AI surfacing in class action litigation, a growing patchwork of state laws, and cyber insurance that may not cover AI claims the way companies assume. AI hallucinations are inventing fake case law in real filings — a plot twist even Kesey wouldn't have tried.

You can only govern what you can see

The takeaway is simple: you can only govern, budget for, and protect what you can actually see. Nurse Ratched at least kept a chart on every patient. Most Security teams can't say the same about their agents. Let's go a little deeper — into what it looks like when the asylum isn't being run operationally at all.

A cybersecurity startup went from stealth to a $640 million valuation in four months. Onyx Security closed a $113 million Series B last week, led by Bessemer Venture Partners. The company builds a control layer for AI agents — permissions, activity monitoring, and remediation when an agent's reasoning goes somewhere it shouldn't. Anthropic integrated Onyx's technology into its enterprise offering in June. Revenue quadrupled since the company left stealth. Basically, someone finally hired a head nurse.

Read the Series B round size correctly, and it says something about hiring, not just funding. Agent governance is becoming a standing line item in every enterprise Security budget. That means a new job exists now: engineers who monitor an AI agent's decision path the way AppSec engineers monitor code, and who understand non-human identity the way legacy IAM engineers understand human identity. Call it the orderly role nobody put in the job posting.

Most security teams don't have that person yet

Most Security teams don't have that person yet — and most companies haven't even considered what AI tooling has already unleashed inside their own environment. It's the Jurassic Park problem in miniature: everyone was so preoccupied with whether they could deploy an agent, nobody stopped to ask who's supposed to watch it after the demo ends. That's the gap the next wave of governance vendors is betting will get funded before it gets fixed the hard way.

Ask most security teams who owns agent oversight today and you'll get the same blank stare the Bobs got in Office Space when they asked what, exactly, someone does here. Somebody's job is agent governance. Nobody's told them yet. Meanwhile Skynet never needed a coup — it just needed nobody watching the deployment logs.

Every major technology wave eventually creates a governance layer — cloud led to CSPM and CNAPP, APIs led to API Security, and autonomous AI agents are now driving the need for AI agent governance. As enterprises move beyond copilots to autonomous execution, capabilities like identity, authorization, policy enforcement, observability, and remediation will increasingly become foundational platform services rather than application-specific features. The long-term winners may be the ones who make enterprise AI trustworthy and governable at scale — the ones running the ward instead of getting run by it.

Agent governance is a budget line item. Agent ROI has yet to be found.

We still need to talk about the mental health of agents too. Somewhere, Nurse Ratched is smiling — she finally got her whole staff automated.