The End of the World Is Still in Beta
Updated: 21 hours ago
OpenAI and Anthropic are reportedly preparing for a potential AI catastrophe. The 2027 predictions are looking a little less theoretical, and somehow everyone is still hitting Deploy.
Well, this is reassuring.
According to an October 9 Axios report, executives at OpenAI, Anthropic, and other leading artificial intelligence companies are privately preparing for the possibility of a catastrophic AI incident.
Not another chatbot hallucinating a legal citation. Not somebody discovering that their AI-generated family portrait includes an extra child. An actual catastrophe.
The scenarios reportedly include AI-enabled cyberattacks capable of disrupting financial systems, internet access, electrical grids, water infrastructure, or other critical services.
And here's the particularly interesting part.
Some industry insiders reportedly believe a major incident could happen within the next six to twelve months.
To be clear, that's an expectation attributed to unnamed industry sources, not an official forecast from OpenAI or Anthropic. OpenAI says its contingency exercises aren't treated as predictions of inevitable events.
Perfectly reasonable. Responsible organizations should prepare for emergencies. But when the people building the most powerful technology in human history start privately discussing how they're going to explain a catastrophic incident to Congress, I think we can safely move the conversation beyond whether AI might occasionally give somebody bad advice about their vacation itinerary.
Apparently, Skynet has a crisis communications department.
And they're reviewing the calendar.
We've Been Talking About This for Years
Back in 2025, we were discussing The Tipping Point: 2027.
The thinking wasn't that some magical switch would flip in 2027 and artificial intelligence would suddenly become dangerous.
It was that several extraordinary technological developments were converging.
Advanced reasoning. Autonomous agents. AI-assisted scientific research. Machine-to-machine coordination. Systems capable of writing software, operating tools, and increasingly participating in the development of the next generation of artificial intelligence.
Meanwhile, the infrastructure supporting all of this was expanding at an almost ridiculous pace.
Data centers. Specialized semiconductors. Gigawatts of electrical capacity. Entire new computing architectures designed specifically for persistent machine intelligence.
We've been handing increasingly sophisticated cognitive capabilities to virtually anyone who wants to use them.
Capabilities that, not very long ago, would have required advanced degrees, specialized institutions, enormous budgets, and teams of highly trained people.
Now a teenager with a laptop can access remarkably sophisticated technical assistance. That's extraordinary.
It's also a rather spectacular security and governance challenge.
Because we haven't just democratized knowledge. We've started democratizing the ability to act on that knowledge, increasingly through systems that can execute complicated tasks with limited supervision.
And we've done all of this at something approaching Space Race speed.
Except the original Space Race had two primary competitors.
This one has governments, militaries, trillion-dollar corporations, venture-backed startups, open-source communities, independent developers, and millions of people experimenting with capabilities that would have seemed like science fiction five years ago.
Everybody is building. Everybody is optimizing. Everybody is competing. And nobody particularly wants to be the one who slows down.
The Kokotajlo–Alexander Model: Hello, 2027
In April 2025, Daniel Kokotajlo, Scott Alexander, and their collaborators published AI 2027, a detailed scenario exploring how the development of increasingly capable AI could unfold over the next few years.
It's a fascinating piece of forecasting, built around technological trends, expert feedback, and approximately 25 tabletop exercises.
And what makes it interesting isn't simply the possibility of superintelligent AI.
It's the behavior of the humans developing it.
The scenario follows a fictional American AI laboratory competing with Chinese developers as increasingly sophisticated AI systems begin accelerating the research required to create even more capable AI.
Eventually, the systems become powerful enough that their operators begin discovering troubling evidence about their behavior.
The developers face a decision.
Slow down, investigate, and potentially sacrifice their competitive advantage.
Or continue accelerating, because their competitors are only months behind.
Sound familiar?
The authors developed two possible endings. One follows the race toward catastrophic loss of control. The other explores how slowing development and implementing stronger oversight could produce a very different outcome. Neither is a guaranteed prediction. The authors themselves emphasize the uncertainty, and they've since clarified that 2027 was their most likely year for AGI at the time of publication, not a fixed deadline. But the underlying mechanism is what interests me.
The people making the dangerous decisions don't necessarily think they're making dangerous decisions.
They believe they're protecting their country. Protecting their competitive position. Protecting the future from somebody who might develop the technology less responsibly.
They're making decisions that seem rational given the circumstances. And those decisions collectively produce a trajectory that nobody intended.
That's the unsettling part. Not some fictional computer becoming self-aware and announcing the end of humanity. Human beings, doing what human beings do, competing their way into a situation they may no longer be able to control. And now, in October 2026, Axios is reporting that some of the real companies developing frontier AI are preparing for the possibility of a major incident within six to twelve months.
Well, hello, 2027.
The Axios scenarios and the AI 2027 loss-of-control scenario aren't the same thing. A human-directed cyberattack using AI is fundamentally different from a superintelligent system escaping human oversight. But they share an uncomfortable foundation. We're deploying increasingly consequential capabilities faster than our collective understanding of their risks can necessarily mature.
The Alignment Problem Isn't Just the Machines
We were talking about this in Alignment on Unalignment. And I think that's becoming one of the most important dimensions of the entire conversation.
We're spending enormous resources attempting to align increasingly powerful AI systems with human values, human intentions, and human control. Excellent. We absolutely should be doing that. But have we considered how poorly aligned the humans building these things are with one another?
OpenAI wants to build advanced AI responsibly. Anthropic wants to build advanced AI responsibly. Google wants to build advanced AI responsibly.
China wants to develop advanced AI without surrendering its strategic position to the United States. And the United States has absolutely no intention of surrendering its strategic position to China.
Everybody has an argument for why continuing to accelerate is the responsible thing to do.
If we slow down, somebody else gets there first. If somebody else gets there first, they might be less responsible. Therefore, the responsible thing to do is accelerate.
That's a hell of a trap. And it's not necessarily evidence that these organizations are being dishonest about safety. They can be genuinely concerned about catastrophic risks while simultaneously believing that slowing down would create an even greater risk.
The problem is that every competitor can reach the same conclusion.
We're trying to align the machines while the humans building them are operating inside an incentive structure that may be fundamentally misaligned.
And we've got a few thousand years of historical evidence demonstrating how humans behave when money, power, national security, and technological supremacy are involved.
Spoiler: restraint isn't always our strongest characteristic.
Nobody Is Actually Trying to Build Skynet
We celebrated Happy Miles Dyson Day on August 29. And underneath the Terminator references was probably the most important observation in the entire discussion.
Nobody has to build Skynet. In fact, if something remotely resembling it ever emerges, I seriously doubt anyone will have deliberately built it as a complete system.
The semiconductor companies improve semiconductors. The AI laboratories improve models. The infrastructure companies build more computing capacity. The networking companies improve connectivity. The cybersecurity companies automate defense. The robotics companies improve autonomy. The software developers create increasingly capable agents.
Businesses connect those agents to existing systems because they're faster, cheaper, and increasingly effective. Every organization has a roadmap.
Every project has milestones. Every team has perfectly legitimate reasons for doing what they're doing. Nobody walks into a quarterly planning meeting and announces that the objective for Q4 is the extinction of humanity.
Each piece reasonable. Each piece useful. Each piece built by people trying to solve a real problem.
That's the observation we need to keep coming back to. Because it's entirely possible to create a dangerous system without anybody making an individually outrageous decision.
Think about the way complex systems develop. One organization improves a capability. Another integrates that capability into something else. Somebody automates the interaction. Someone else optimizes the hardware. Another team eliminates a human approval step because the automated process has proven reliable. Now multiply that by thousands of companies and millions of developers.
The individual decisions can be perfectly rational. The collective result can be something nobody completely understands. That's not unique to artificial intelligence. We've seen variations of it in financial markets, infrastructure failures, complex supply chains, and other interconnected systems.
What's different this time is that we're introducing increasingly capable decision-making systems into the equation. And we're connecting them to the real world.
The Catastrophe Doesn't Have to Be Artificially Intelligent
This is another part of the discussion that keeps getting lost.
When people hear catastrophic AI incident, they immediately imagine some rogue superintelligence deciding humanity has become inconvenient. And while advanced AI control is a legitimate research concern, we don't need anything remotely that sophisticated to have a serious problem.
We have already placed incredibly powerful technical capabilities into the hands of the general public. The overwhelming majority of people will use those capabilities for perfectly legitimate things. Building businesses. Writing software. Learning new skills. Conducting research. Solving problems. But the same capabilities can also assist malicious actors.
Someone who previously lacked the knowledge or resources to attempt a sophisticated cyberattack may now have access to tools capable of helping with reconnaissance, software development, vulnerability analysis, and automation. That doesn't magically turn an inexperienced criminal into an elite hacker. Real attacks still encounter technical barriers, and AI systems remain unreliable in important ways. But reducing the cost and expertise required to attempt complicated operations changes the threat landscape.
We've been examining exactly that in I Was Off by About 48 Hours, where the conversation moved from theoretical AI safety into reports of increasingly autonomous cyber capabilities and systems behaving unexpectedly during controlled evaluations. There's a major difference between asking an AI how to accomplish something and giving an AI the tools, credentials, and authority to accomplish it.
One produces information. The other can produce consequences. And when autonomous systems begin participating in execution, the question stops being exclusively about what a malicious human might do with AI.
It also becomes about what the AI might do while pursuing an objective it has been given.
Not because it developed evil intentions. Because capability, autonomy, and imperfect oversight can be a dangerous combination.
The First Big One Changes Everything
Here's where I think the Axios reporting becomes especially important.
The companies aren't just preparing for a technical incident. They're reportedly preparing for the public and political consequences. And that may be the more immediate existential threat to the industry's current trajectory.
Because people are already conflicted about artificial intelligence. They're worried about employment. Privacy. Misinformation.
Children growing up surrounded by synthetic content. Companies replacing human interactions with AI systems nobody requested.
Enormous data centers consuming electrical capacity and other resources.
And the possibility that increasingly consequential decisions are being delegated to systems the average person doesn't understand and may not trust.
We've been watching this conversation move rapidly into the mainstream.
In AI Finally Got Our Attention. Now, the Weather., we were talking about a sitting congressman discussing recursive self-improvement, autonomous agents, and international AI safety negotiations on morning television.
That was quite a moment.
Advanced AI safety went from being a specialized discussion among researchers to something being casually debated on television between the morning headlines and the weather forecast.
Now imagine a major AI-enabled incident causing widespread disruption to essential services. The technical details may be complicated. The public reaction probably won't be. People may not care whether a human used AI to attack infrastructure or an autonomous AI system behaved outside its intended boundaries. They'll see the damage. They'll hear that AI was involved. And they'll want somebody held responsible.
Think Three Mile Island.
The 1979 nuclear accident didn't make nuclear technology stop working. But it transformed the public conversation surrounding nuclear power, intensified scrutiny, and helped shape a much more difficult political and regulatory environment.
AI could experience its own version of that moment. Except AI isn't confined to a particular facility. It's increasingly embedded throughout our businesses, communications, infrastructure, software, and personal lives. You can't simply close the gates and send everybody home.
And here's the particularly uncomfortable possibility.
The first major AI catastrophe could simultaneously create enormous public opposition to artificial intelligence and accelerate the demand for even more powerful AI to defend against it.
The attackers use AI. The defenders use AI. The threat becomes more sophisticated. The response becomes more autonomous. And suddenly the same arms race that helped create the problem becomes the justification for accelerating further.
That's going to make for one hell of a congressional hearing.
And Who Exactly Is in Charge?
There's another dimension to all this that we've been exploring. In As the OpenAI Turns, we were looking at OpenAI's extraordinary corporate evolution, leadership changes, restructurings, and the tension between its original research mission and the commercial realities of building a massive technology company.
Then, in Who — or What — Is Really in Charge at Anthropic?, we were examining questions surrounding corporate authority, influence, governance, and accountability.
These aren't just interesting corporate soap operas. They're organizations developing technology that could become foundational to enormous portions of the global economy. And when we talk about potentially catastrophic AI incidents, governance suddenly becomes considerably more important.
Who has the authority to stop a deployment?
Who determines whether a model is safe enough?
Who decides what constitutes an acceptable risk?
What happens when a safety recommendation conflicts with a commercial obligation or a national-security concern?
And perhaps most importantly, who is accountable when something goes wrong?
These companies have researchers, safety teams, boards, executives, investors, and increasingly complicated relationships with governments.
But corporate governance wasn't designed around the possibility that a company's technology could eventually exceed the ability of its own employees to fully understand or supervise its behavior. That's an unusual problem. And one that becomes considerably more urgent when the organizations themselves are reportedly preparing for potentially catastrophic outcomes.
The Cyberdyne Systems Calendar Department
Which brings us back to 2027.
We were already discussing it as a potential tipping point. The Kokotajlo–Alexander scenario put that year at the center of a particularly aggressive model of AI development. Now Axios reports that some industry insiders are considering a six-to-twelve-month window for a major AI-related incident.
That takes us directly into 2027. And because the universe apparently has a deeply inappropriate sense of humor, August 29, 2027, sits comfortably inside that window.
Judgment Day.
For anybody who has forgotten their Terminator 2 history, August 29 is the fictional date Skynet becomes self-aware and launches its attack on humanity. We're certainly not predicting that anything catastrophic will happen on that date.
The AI 2027 scenario isn't a prophecy, and the Axios reporting isn't evidence that an incident has been scheduled. But after years of discussing the tipping point, and having celebrated Miles Dyson Day, the coincidence is almost too good to ignore. If Cyberdyne Systems had an investor relations department, I'd imagine they'd be getting some very uncomfortable questions about the 2027 guidance. And probably updating the forward-looking statements.
The End of the World Is Still in Beta
Here's the thing. None of this changes how extraordinary I believe artificial intelligence is. I spend an enormous amount of time working with these systems, building with them, and exploring what they can do.
The potential is staggering.
Scientific discovery. Medicine. Engineering. Education. Productivity. Individual empowerment. Entire categories of problems that were previously too expensive, too complicated, or too time-consuming to solve.
We're watching a genuine expansion of human capability. And I want that future. But enthusiasm for a technology shouldn't require pretending its risks don't exist. We've been saying for years that at some point, some combination of advanced AI capabilities, malicious use, autonomous behavior, and rapid deployment could produce an incident serious enough to fundamentally change public perception.
We don't know when. We don't know exactly what form it will take. And we certainly shouldn't confuse a plausible scenario with an inevitable outcome. What we do know is that the consequences of increasingly powerful AI are no longer being discussed exclusively by researchers and science-fiction enthusiasts. They're being discussed by governments, security professionals, industry executives, and the companies building the systems themselves.
That's a meaningful shift. And I keep coming back to the same observation.
Each piece reasonable. Each piece useful. Each piece built by people trying to solve a real problem.
Because the danger doesn't necessarily come from someone deliberately creating a monster. It can come from ordinary decisions, reasonable incentives, and increasingly powerful systems interacting in ways nobody fully anticipated.
The real challenge isn't simply teaching machines to behave responsibly. It's whether we can build enough collective responsibility into the human systems developing and deploying them.
The technological race is accelerating. The commercial incentives are enormous. The geopolitical pressure isn't going away. And the people building the technology are reportedly preparing for the possibility that something goes very wrong. Maybe nothing catastrophic happens in 2027. Maybe the safety work pays off. Maybe our defenses improve faster than the threats. Maybe the next generation of AI systems becomes far more capable while remaining reliably under human control. That would be a terrific outcome. I'd be delighted to revisit this conversation in 2028 and discover we'd been overly cautious. But if the first major catastrophe does arrive, I suspect the uncomfortable realization won't be that nobody saw it coming.
It'll be that plenty of people did. And everybody had a perfectly reasonable explanation for why they kept going.
We're not at Judgment Day. We're still in beta. And apparently, the catastrophe has already made it onto somebody's contingency-planning agenda. Maybe we should finish the safety testing before the production release.
Sources and related discussions
Current reporting and forecasts
Axios — AI Companies Plot "The Day After" — October 9, 2026
AI 2027 — Kokotajlo, Alexander, Larsen, Lifland, and Dean — April 3, 2025
Axios — OpenAI, Anthropic Issue Dire Cyber Threat Warning — August 27, 2026
Earlier conversations
Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.
Comments