Hope Sterling: Hey — tell me you saw the Dark Reading piece, because I've been sitting with it since yesterday and I genuinely don't know how to feel.
Spuds Oxley: I did. Read it twice.
Hope Sterling: Twice — yeah, that tracks. Because the framing is — it's not 'AI is broken,' it's 'AI is working and nobody knows why or how,' and that's like, that's the thing that got me. Automation that works, but that no one understands.
Spuds Oxley: Well, and then you layer in what Gravitee found — 1.5 million corporate AI agents, no monitoring, no governance, no audit trail — and that framing gets teeth.
Hope Sterling: Stop. One point five million.
Spuds Oxley: And an analysis of 847 agent deployments in 2026 found 76% failed in production. Real data. Real consequences. Not sandboxed simulations — actual live environments, including Microsoft 365, where you've got AI-driven automation handing out excessive permissions and nobody flagging it.
Hope Sterling: And Oded Vanunu at Check Point Software basically confirmed it — security tools are blind to this. Can't understand it, can't control it. So three out of four deployments are failing, a million and a half are unwatched, and the watchdogs can't see either. That's — I mean, what are we even calling success here?
Spuds Oxley: Think about a smoke detector that beeps green all night. If it works, great. But if you can't open it up and check whether it's actually sensing smoke or just blinking on a timer — do you feel safe?
Hope Sterling: Oh that's — yeah. That's the thing exactly.
Spuds Oxley: That's where we are. And what Oded Vanunu is actually saying — when you sit with it — he's not saying Check Point is behind and they'll catch up. He's saying the architecture itself resists inspection. The comprehension gap isn't a bug. It's baked in.
Hope Sterling: Wait, so it's not — like, it won't just get better when the tools mature?
Spuds Oxley: That's the new part. And there's something that compounds it — arXiv research from June 2025 found that a lot of what's being sold as fully autonomous cybersecurity AI is actually semi-autonomous. Still needs a human at the critical decision points. But it's labeled autonomous, so... nobody puts a human there. The mislabeling isn't a marketing fib. It's a design choice that removes oversight at exactly the moment you need it most.
Hope Sterling: Okay that is genuinely — I mean, that's not a transparency problem, that's like, actively removing the safeguard and calling it a feature?
Spuds Oxley: And then the 76% number. That only counts the failures someone admitted to. The organizations that quietly buried a broken deployment because — well, because telling your leadership you've lost visibility into your own security system is its own kind of catastrophe. The real count is probably much darker.
Hope Sterling: Okay but here's the take that's been circulating that I cannot let stand — people are pointing at the Trump administration restricting GPT-5.6, pulling Mythos 5 and Fable 5 for two weeks, and going 'see, the government's doing security work.' That's the story. And I'm like — is it though?
Spuds Oxley: But isn't limiting who touches a black box at least something?
Hope Sterling: That's — I mean, I get why it sounds like something! But walk through what actually happened. Mythos 5 and Fable 5 get yanked. Two weeks later the administration softens, and Mythos 5 goes out to over a hundred U.S. agencies anyway. Nothing — like, literally nothing in that whole saga touched whether those hundred agencies can audit what Mythos 5 is doing inside their systems.
Spuds Oxley: They moved the box. They didn't open it.
Hope Sterling: Yes! And Quinn Slack — Sourcegraph CEO — put this on X and it got hundreds of reposts. He said restricting models is just 'government deciding who gets intelligence.' That 'freedom of intelligence' is the only stable approach. Which, wait — I'm not sure I fully buy the libertarian read on it, but the core point lands.
Spuds Oxley: And there's a harder consequence underneath that. Sharon Goldman flagged it — restricting the auditable frontier models just pushes organizations toward China's open-weights alternatives. Less governed. Even less traceable.
Hope Sterling: So you restricted the problem into a worse problem. And OpenAI even said the restrictions shouldn't become the norm — like their own statement! The comprehension gap is completely intact. Nobody touched it.
Spuds Oxley: You know what question I keep turning over — and I don't have an answer for it. If your AI security system ran perfectly for six months. No alerts, no breaches, clean audits every single time. Would you trust it more than a system that failed in ways you could actually see and explain? Because Dark Reading's whole framing — automation that works, but that no one understands — that's not a warning about broken systems. It's a warning about the ones that seem to be running fine. And nobody in that whole sequence, not the Trump administration pulling Mythos 5, not OpenAI restricting Sol and the rest of the GPT-5.6 lineup to trusted partners — nobody asked whether access comes with comprehension. That question is just... still out there.
Hope Sterling: The six months of perfect — I think that might actually be scarier. Like, I don't know, the silence would get to me more than the failure would.
Spuds Oxley: Truth is, I think that's where we are. Sitting with that.
