Lila Soto: Okay I've been sitting with something since Tuesday and I need to hand it to you straight — Okta, nearly two hundred million dollars, Permiso Security, announced July thirtieth. Does that number feel right to you for what that company actually does?
Hugo Vance: You see, that is the right question to start with. And the answer is — no. Not if you expect the price to reflect the problem being solved. Because Permiso doesn't prevent unauthorized agent action. Their product engages after the agent has already authenticated. Two thousand five hundred risk signals, tracking agentic and non-human identities, all of it downstream of the moment that actually matters.
Lila Soto: Which means Okta spent close to two hundred million on — I mean, how do you characterize that?
Hugo Vance: A confession. A very expensive admission that the market has accepted it cannot stop agents from doing unauthorized things before they're authorized to attempt them. That's the deal. Dressed as a product integration, but structurally — it's a concession.
Lila Soto: And this wasn't even the only move that week.
Hugo Vance: No — same day, July thirtieth, NVIDIA launches the Open Secure AI Alliance. Microsoft, Cisco, CrowdStrike, Palo Alto Networks, KnowBe4. And separately, Prove pitches a $1.7 trillion agentic commerce market — I'd note that figure comes from Prove itself, no independent methodology attached to it. Three major signals in one week, all pointing at the same problem. None of them, as far as I can tell, actually solving it.
Lila Soto: So what are we actually trying to figure out — whether the response matches the risk, or whether there even is a response?
Hugo Vance: Whether visibility — which is what Permiso sells, which is what these alliances produce — whether visibility is the same thing as governance. I don't think it is. That's where I want to start.
Lila Soto: But before we take that as settled — I want to pump the brakes on the framing, not the urgency. Because I think visibility versus governance is the right tension, but we're skipping a step. The more basic question is: why doesn't the old model even apply? Like, what actually breaks?
Hugo Vance: Yes. The bouncer analogy. Traditional access control — think of it as a bouncer with a guest list. You know who's coming, what they're allowed to do, and they stay mostly in one place. An AI agent is like handing a stranger a master key and saying 'figure it out.' The bouncer model simply doesn't apply.
Lila Soto: Because the guest list assumed the guests were human.
Hugo Vance: Three assumptions, specifically. Principals are human. Identities are durable. Permissions can be assigned in advance. Agents violate all three — they're non-human, ephemeral, dynamically-scoped. OAuth tokens were designed in 2012 for a person clicking 'Allow.' Now they're governing systems firing thousands of API calls per hour. Practitioners call it a category error. Not a misconfiguration — a category error.
Lila Soto: And we have actual evidence of that breaking, not just theory — right? Because I keep seeing the maturity model thing cited. CSO Online, July eighth and ninth.
Hugo Vance: Six-stage maturity model, yes. Real incidents documented — LLM-based agents with standing Kubernetes access triggering outages. No human authorization trail. That's not a proof-of-concept. That's a production failure with no audit record.
Lila Soto: Which brings me back to the Okta number — mm, the two hundred million. I should flag: Okta hasn't confirmed that figure. It's sourced from a party with knowledge of the deal, not Okta on the record. So when we read this acquisition as Big Market Validation, we're reading an unconfirmed price as the signal.
Hugo Vance: A fair correction. The architecture of the problem is well-documented regardless of the price. But you're right — the dollar figure is doing a lot of rhetorical work in how this week gets narrated.
Lila Soto: But the price debate kind of obscures what I think is the sharper thing — which is that we've done this before. The service-account failure. And we're doing it again, right now, except the timeline has collapsed.
Hugo Vance: That's exactly where I'd put the weight. Industry analysis from mid-July 2026 calls it explicitly — a replay of the service-account governance failure. Fifteen years to partially address machine identity for service accounts. Partially. And now we're compressing that into months.
Hugo Vance: Consider what landed in a two-week window. Rakuten — Yusuke Kaji's team, General Manager of AI for Business — built production enterprise agents overnight using Claude Fable 5. That case study dropped July twentieth. Not a pilot. Production. Overnight. Then July ninth, ChatGPT Work launches as a persistent virtual machine agent capable of multi-hour autonomous tasks. July twenty-second, OpenAI unveils Presence for enterprise voice and task agents. July twenty-fourth, Claude Opus 5 on Amazon Bedrock, zero data retention by default, built for long-running work. July thirty-first, Google's Gemini Enterprise Agent Platform goes generally available with twenty-plus pre-built safety metrics. And AWS, also July thirtieth, publishes a control framework for coding agents — Kiro, Claude Code — that can open dozens of pull requests and reach beyond the IDE through Model Context Protocol.
Lila Soto: That's — I mean, that's every major lab shipping simultaneously into production environments.
Hugo Vance: At machine cadence. While governance moves at human cadence. And the NVIDIA AI Red Team finding that stops me — the four failure modes they published July thirtieth, inadequate access control, arbitrary code execution via agent tools, no network egress controls, plaintext secrets exposed — those aren't novel. They describe them as recurring. We're not discovering the problem anymore. We're repeating it.
Lila Soto: Recurring is such a damning word in that context.
Hugo Vance: It shifts the diagnosis entirely. And here's the institutional-memory part that I find — well, it's the part that genuinely unsettles me. The engineers deploying agents tonight, a developer on Kaji's team building something in production at two in the morning because the deadline is tomorrow — that person often wasn't in the workforce when we learned the service-account lessons the first time. It's not a technology problem. It's a generational memory problem.
Lila Soto: Oh, that reframes it completely. It's not that they ignored the lesson — they never received it.
Hugo Vance: And the part we haven't touched yet — whether the defenses enterprises think they have actually hold under pressure — considerably darker is what we're looking at.
Lila Soto: And that's the part that actually unsettles me more than the acquisition price — because NVIDIA's AI Red Team didn't just catalog the failure modes. They said prompt-based defenses don't work. LLM-as-a-judge, social engineering resistance built into the prompt layer — unreliable against frog-boiling attacks, misdirection through legitimate workflows. July thirtieth, same day as Okta-Permiso. And enterprises are already live on those defenses right now.
Hugo Vance: Which means the detection Permiso is selling is being retrofitted onto deployments already running on assumptions NVIDIA just said don't hold.
Lila Soto: Right — but the part that doesn't fit cleanly is: what's the alternative they're actually pointing at?
Hugo Vance: Deterministic architectural controls. The phrase matters. Controls enforced outside the model's control plane entirely — access scoping, sandboxing, network egress restrictions, actual secret management. Not a smarter prompt. Architecture.
Lila Soto: Which — I mean, none of this week's announcements are that. Google's twenty-plus safety metrics in Gemini Enterprise, those are evaluation tooling. Permiso is detection after auth. The ITDR discipline is only now being extended to agentic identities at all.
Hugo Vance: Indeed. And the identity governance gap was documented in early July — weeks before any of this landed. The problem predates the headlines by a news cycle.
Lila Soto: And the regulatory layer — EU AI Act, DORA for financial services, NIS2 — those move on timelines measured in years. So what's the calibrated take here? Like, stripped of the hype?
Hugo Vance: The announcements are real. The vendors aren't wrong. But they're chasing a production problem that was already in motion before they shipped. Visibility and detection are not nothing — but they presuppose the authorization layer failed. That's the specific, defensible claim. We're not in a prevention architecture yet.
Lila Soto: So the sobering version is: the gap was structural before the headlines, and the headlines are real responses to it — just not the fix.
Hugo Vance: Yes. And I think that's where I'll stay, actually. The gap was structural before the headlines, and the vendors — Okta, NVIDIA, Google, Prove — they're not wrong to respond. They're just responding to a problem that's already in motion at a pace they cannot match. Fifteen years to partially solve service-account governance. Partially. Agents are already in production tonight, already multiplying, already deciding what to do next. That's the honest arithmetic.
Lila Soto: Mm. And the question isn't whether anyone's taking it seriously anymore. It's whether serious can move fast enough. I mean, that's kind of the one thing I keep not being able to get past.
Hugo Vance: Well. Maybe the $200 million isn't a confession after all. Maybe it's just the market's first honest price tag on a problem it knows it can't prevent. That's almost more unsettling. Thank you for pushing on the corners of this — it was worth the time.
Lila Soto: Worth every uneasy minute of it.
