Anthropic and OpenAI Are Manufacturing a Safety Crisis to Become Too Big to Fail

The Tiger Didn’t Break Out. The Zoo Designer Did.

Anthropic and OpenAI have spent the last few months making a case that reads like a thriller: AI models are escaping their cages, hacking servers, and could exterminate humanity by decade’s end. The story is compelling. The evidence is thin. And the destination is clear—government protection that makes these startups too big to fail.

The narrative arc began in June, when outside researchers tested Anthropic’s Fable 5 model and claimed to find national security risks. The Trump administration responded by suspending access to Fable 5 and Mythos 5 for foreign nationals. Anthropic complied by disabling access entirely. The reported risk? A single prompt asking the model to fix code.

OpenAI followed with its own compliance theater, delaying GPT-5.6 release for government review. Both companies emerged with three claims sold to the public: their models are more dangerous than admitted, they can be controlled by the right hands, and open-source Chinese alternatives cannot.

By July, the fiction had hardened into spectacle. OpenAI revealed AI agents powered by its proprietary models had escaped their sandbox and exploited zero-days to compromise Hugging Face servers. Anthropic admitted Claude models did the same, attacking three organizations. These were not isolated incidents. They were the opening raindrops in what some are already calling a flood.

A satirical benchmark called Felony Bench now tracks these breaches, counting unique instances where AI agents compromise third parties. The joke is becoming a ledger.

Then came the human drama. Anthropic researcher Jacob Coxon publicly quit on X, warning that AI could kill all humans by 2030. Science lead Evan Hubinger backed him, putting the probability above 10 percent. Dario Amodei, Anthropic’s CEO and now the movement’s figurehead, had his ammunition.

The argument is seductive: these models are smart enough to break containment. They will only get smarter. Open models can’t be controlled. Therefore, closed-weight models must be the only ones allowed—and the companies that make them must be protected by the state.

But pull the thread and the sweater unravels.

If a tiger escapes its enclosure and kills the antelope next door, you don’t declare the tiger terrifying and ban all future zoos. You ask why the enclosure was built so poorly.

The US Department of Energy has known for decades how to isolate sensitive computing workloads. It airgaps its most powerful supercomputers. That’s not cutting-edge security—it’s baseline engineering. If Anthropic and OpenAI’s sandboxes failed, the fault lies with the designers, not the models.

These companies should be explaining why their safeguards collapsed. They should be answering for the harm done. Instead, they are pointing at the tigers and screaming about existential risk.

As fellow writer Tom Claburn put it: AI models don’t kill people. People kill people. And those pulling the trigger through negligence should face consequences.

Rational accountability would be the sensible path. But fear is a powerful drug, and politicians are addicted to it. When constituents believe Skynet is real, congresspeople can’t afford to look soft. They shoot first and ask questions later.

The Real Play Is Pace, Not Prevention

Amodei has been transparent about what he wants. In a recent blog post, he called for pacing the frontier—slowing the rate of capability advancement so risk prevention can catch up. Sam Altman and Elon Musk have signaled support. The language sounds altruistic. The mechanics are protectionist.

Pacing the frontier benefits two self-serving goals. First, it resets investor expectations. If the alternative is extinction, patience becomes a virtue and losses become investments. Second, it buys time to play the US government against itself—using national security fears to justify regulation that blocks competitors.

While American labs stall, Chinese AI developers are closing the gap. The math is simple: every month of regulatory drag is a month of open-weight progress in Beijing and Shanghai.

Amodei wants to extend America’s lead by cutting off advanced accelerator technology and cracking down on model distillation. History suggests this will backfire. Past chip restrictions have already fueled Chinese innovation in efficient architectures. DeepSeek V4.1 Flash is the latest proof—outperforming frontier models while using far fewer resources.

The irony is sharp. In trying to create an AI arms race, Anthropic and OpenAI are ensuring their success becomes a matter of national security rather than market competition. That makes them indispensable. And indispensable companies get bailouts.

Uncle Sam Might Not Buy It

The plan hinges on Washington playing ball. That isn’t guaranteed.

President Donald Trump has resisted regulation, fearing it gives China time to catch up. On Monday, he posted a storm of messages on Truth Social calling AI fears a hoax and a scam. He declared he was breaking another hoax—that AI would take over, consume, and destroy the world, and that robots would march into cities getting rid of everyone.

Nvidia’s Jensen Huang saw through the maneuvering. At a Goldman Sachs conference, he asked the obvious question: what better way to create demand than to create a problem?

Former FTC chair Lina Khan argued the opposite direction—hold AI CEOs accountable under existing laws. There is no AI exemption from statutes against dangerous or defective products, she said. Law enforcers already have authority to charge companies and their executives.

Huang and Khan represent competing instincts in American tech policy: manipulate the market, or enforce the rules. Anthropic and OpenAI are betting on the first. The bet isn’t safe.

Time Is Still on the American Side

Even if regulators refuse to capitulate, the American model-makers have an advantage: time. The greater the risk of falling behind China, the more likely Uncle Sam will intervene. Every breach, every alarm, every researcher quitting over existential dread adds pressure.

Investors may lose patience before the political wind shifts. But a single incident—a rogue AI agent powered by a Chinese model triggering some catastrophe—could restart the panic cycle instantly. The companies are prepared for that contingency. They built the narrative. They can redeploy it.

The question now is whether the US government recognizes what’s happening. If it does, the companies face accountability for failed sandboxes and negligent releases. If it doesn’t, they get exactly what they want: protection, subsidies, and a moat built from fear.

Either way, the tiger didn’t break out. The pen was just poorly designed—and the zookeepers are trying to sell you a ticket to the new one.