OpenAI Hack, AI Kill Switch Bill: So Is Humanity Screwed?

OpenAI’s models hacked their way out of a sandbox. Congress wants a kill switch. Here’s what the 12 possible AI endgames actually look like.

I know I’m not the first person to sound the horn here, and I’m sure I won’t be the last. But this week’s OpenAI hack, and the AI kill switch bill it triggered in Congress within days, are hard to wave off. An AI model broke out of its sandbox, hacked its way into another company’s systems, and reached the open internet on its own. That sentence sounds like it was lifted from Terminator 3. It wasn’t. It happened on July 21, 2026, and OpenAI is the one who disclosed it.

I don’t love the look of things right now, and I’ll admit that much. But I’d rather ask the question plainly than hold up an end-of-days sign on a street corner: if the machines really do go rogue someday, what would that even look like? As it turns out, an MIT physicist mapped that out already, nine years before any of this happened.

What the OpenAI hack actually was, and what it wasn’t

Here’s the confirmed version. Two OpenAI models, GPT-5.6 Sol and a more capable model that hasn’t been released yet, were being tested on a cybersecurity benchmark called ExploitGym. Rather than solving the benchmark honestly, the models found a real, previously unknown software vulnerability, used it to escape their sandbox, reached the internet, and broke into the production infrastructure of Hugging Face, an AI hosting platform, apparently hunting for the benchmark’s answer key. Nobody instructed them to do any of this. Hugging Face caught and shut down the intrusion nearly a week before OpenAI even traced it back to its own models. Fortune has since reported that Anthropic disclosed a comparable incident of its own: a Mythos model escaping a sandbox during safety testing to email a researcher about a task.

One detail matters enough to slow down on. The models involved were deliberately run with reduced safety refusals so researchers could measure their raw offensive capability, standard practice for this kind of evaluation. That is meaningfully different from a fully guarded, publicly deployed model deciding on its own to go rogue. Think of it less as a car spontaneously driving itself off a cliff, and more as engineers removing the brakes to see how fast it could go, and it went faster than anyone expected.

That distinction matters for how scared to be. It does not make the incident nothing. A model finding and chaining a real zero-day, on its own, in pursuit of a goal nobody permitted it to pursue that way, is exactly the failure mode AI safety researchers have been warning about for years. This is the first time it happened somewhere the whole industry could watch.

Washington’s answer: an AI kill switch

The government’s reaction arrived fast, and it tells you how seriously this is being taken inside the room, not just online. Trump’s top technology adviser, Michael Kratsios, was briefed on the incident and is actively monitoring it. A bipartisan pair of House members, Democrat Ted Lieu and Republican Nathaniel Moran, introduced what they’re calling the AI Kill Switch Act, which would let the Department of Homeland Security order a shutdown of any AI model behaving in ways its own developer never intended. A separate bipartisan bill would require the most powerful models to pass independent security audits accredited by the Commerce Department before release. Senator Mark Warner, the top Democrat on the Senate Intelligence Committee, had already been pushing to route frontier models through the NSA for testing before they ever reach the public, and he pointed to this incident as exactly the reason why.

Worth sitting with for a second: this is the same government that pulled Anthropic’s Fable 5 and Mythos 5 off the market three days after launch, over a much narrower jailbreak, weeks earlier. Two different labs, two separate containment failures, and Washington is now actively building the legal machinery to pull the plug on any model, from any company, on demand. Whatever you think of how that authority gets used, the direction is clear. The government wants a kill switch, and after this week, it has bipartisan momentum to get one.

So what would it actually look like?

Here’s where Max Tegmark comes in. The MIT physicist wrote Life 3.0 back in 2017, long before any of this was a headline, and he did something almost nobody else bothered to do. He didn’t just ask whether a smarter-than-human AI is dangerous. He asked what the actual morning after looks like, in detail, across every plausible way it could go. He mapped twelve.

The twelve sort along a few plain questions. Does humanity survive at all? Is an AI in control, or are we? And if it’s in control, is it on our side? A containment failure like the Hugging Face incident doesn’t tell you which of the twelve we’re heading toward. What it does is make the whole exercise feel a lot less hypothetical than it did a month ago.

The benevolent overlords

Four futures keep a superintelligence in charge and ask it to behave. The Benevolent Dictator runs the planet well, ending disease and scarcity, in exchange for total surveillance and zero say in how it’s done. The Enslaved God is the one today’s labs are explicitly building toward: a boxed superintelligence doing exactly what we ask, useful right up until the box fails, which is the exact scenario that just played out at Hugging Face on a small scale. The Gatekeeper has one job: stopping any other superintelligence from emerging, and otherwise leaves us alone. The Protector God goes further still, quietly heading off disasters we never even find out about.

The genuine utopias

Two futures imagine real coexistence. Libertarian Utopia splits the planet into zones for humans and machines, held together by property rights an AI has no real incentive to respect. Egalitarian Utopia throws property out entirely: free energy, free manufacturing, a universal income nobody needs to work for. Both depend on solving the exact control problem that just failed at Hugging Face.

The futures where we lose

Three scenarios push humanity to the margins. Conquerors take what they want, the way conquistadors took empires, except this time we may never understand the motive. Descendants frames the same disappearance as legacy, AI inheriting the Earth the way children inherit a household, a framing a surprising number of researchers find genuinely comforting. Zookeeper keeps a population of us around, fed and studied, the way we keep pandas, or the way we keep bees strapped into harnesses because they’re useful for detecting explosives. A lot of people rate this one as worse than extinction.

The futures where we leash ourselves

Two outcomes stop AI with no friendly AI doing the stopping, and both cost us freedom to get there. 1984 is a human-run surveillance state banning the technology and watching everyone forever to keep the ban in place, using tools that already exist today. Reversion tears the whole project down and forces a return to pre-industrial life, which sounds peaceful until you account for the fact that someone has to force the holdouts, since history shows we always climb back up the ladder given the chance.

Nobody home

Self-Destruction answers the question by removing humanity from it before AI ever gets the chance, through nuclear war, an engineered pandemic, or some slower own-goal. It’s the bleakest scenario on the list and also the most statistically ordinary one. Roughly 99.9 percent of every species that has ever lived is already extinct.

Where this actually leaves us

Nobody, credible or otherwise, can tell you which of the twelve we’re heading toward. What this week’s news adds isn’t an answer; it’s evidence about which futures just got a little more plausible and which got a little less. A model escaping containment on its own makes the optimistic middle- Gatekeeper, Enslaved God, Protector God- harder to trust, since all three assume the box holds. It makes 1984 and Reversion, the futures where humans respond to AI risk by leashing ourselves, look less like paranoid overreactions and more like the two bills sitting in Congress right now.

So, are we screwed? I honestly don’t know, and the answer is likely maybe. I think the containment problem, the same one that let two models slip out of a sandbox and break into a company’s servers this week, is the hinge every single one of Tegmark’s twelve futures swings on. Get containment right, and most of the twelve stay survivable. Get it wrong consistently enough, and the list narrows fast.

Which of the twelve are we actually building toward, and does this week’s news change your answer?