The threat was concrete enough that Anthropic didn't release it widely.
When a single AI tool demonstrated it could silently unravel the infrastructure of modern civilisation — power, water, hospitals, finance — the ideology fell away and the vulnerability remained. The Trump administration, which had entered office dismantling AI safety frameworks as ideological excess, reversed course within days of Anthropic's Mythos model revealing what unchecked capability actually looks like in practice. It is a familiar arc in the human story: the rules we dismiss as unnecessary tend to announce their necessity on their own terms. The question now is whether this recognition arrives early enough to shape something durable, or merely reacts to what cannot be ignored.
- Anthropic's Mythos model can exploit operating system and browser vulnerabilities at machine speed and scale, threatening power grids, water systems, hospitals, and financial networks — not theoretically, but demonstrably.
- The Trump administration, which had branded Anthropic 'radical left' and banned it from federal contracts, was forced into a complete policy reversal within days of Mythos becoming known to officials.
- New agreements with Google, Microsoft, and xAI now allow the Commerce Department to conduct pre-deployment evaluations of frontier AI models — the very kind of oversight the administration had aggressively dismantled.
- AI czar David Sacks, the loudest voice against safety regulation, has departed, and the previously frozen relationship between the White House and Anthropic has begun to thaw.
- With over seven hundred billion dollars in AI investment expected this year and a US-China summit looming, the administration faces the unresolved tension between Silicon Valley's race-to-win pressure and the security realities it can no longer dismiss.
When Donald Trump returned to office, his message on artificial intelligence was simple: step aside. The Biden-era safety frameworks were dismantled quickly, recast as ideological overreach. AI czar David Sacks made the posture explicit, and companies that resisted — Anthropic chief among them — were punished. Trump labelled the company radical left, banned it from federal contracts, and cut it off from working with government partners.
Then Anthropic released Claude Mythos Preview, and the calculation changed. The tool could identify and exploit security flaws in operating systems and browsers at speeds no human could match, and it could target the infrastructure modern societies depend on — power grids, water systems, hospitals, financial networks. Anthropic did not release it broadly; instead, roughly forty American companies received access to find and patch their own vulnerabilities before adversaries developed comparable capabilities.
The shock was immediate. Within days, the White House announced pre-deployment evaluation agreements with Google, Microsoft, and xAI, with the Commerce Department's AI standards body gaining access to frontier models before release. The administration that had torn down safety structures was now building new ones. Sacks was gone. The relationship with Anthropic began to quietly recover.
What shifted was not politics but exposure to consequence. Mythos made the risks concrete rather than theoretical, and that distinction proved decisive. The moment also arrived against a larger backdrop: industry leaders from Musk to Amodei to Altman had each, in their own register, warned that AI development had entered genuinely dangerous territory. The US is racing toward agentic and general AI; China, working around chip restrictions, is pursuing efficiency and integration and remains competitive. When Trump meets Xi later this month, AI governance will dominate the agenda — and the irony is pointed. China has championed open-source development and is actively promoting global safety standards. The US, having abdicated that leadership role, may find the most consequential decisions about how the world governs AI made elsewhere.
Whether Trump can hold this new regulatory line against the Silicon Valley billionaires who funded his campaign — and who see AI as a winner-takes-all race attracting more than seven hundred billion dollars this year — is the question that remains unanswered. The belated recognition of risk is real. Whether it produces anything durable is not yet clear.
When Donald Trump returned to office, he came with a clear message about artificial intelligence: get out of the way. The new administration swiftly dismantled the guardrails the Biden White House had put in place, framing AI safety measures as "woke" overreach. Trump's AI czar, David Sacks, made the position explicit: blue states inserting what he called ideological constraints into AI development were the enemy. The White House directed federal agencies to strip out any mention of climate change or diversity concerns from AI risk assessments. Companies that wouldn't play ball faced consequences—when Anthropic refused to let the government use its tools for autonomous weapons control or mass surveillance, Trump branded the company "radical left" and "full of left-wing nutjobs," banning it from federal contracts and from working with any company that did.
Then, last month, Anthropic released something that changed the calculation entirely. The company unveiled Claude Mythos Preview, an AI tool with a capability that cut through all the ideological noise: it could identify and exploit security flaws in operating systems and browsers at speeds and scale no human could match. More alarming still, Mythos could target the infrastructure that keeps modern societies running—power grids, water systems, hospitals, financial networks. The threat was concrete enough that Anthropic didn't release the tool widely. Instead, it gave access to roughly forty American companies so they could find and patch vulnerabilities in their own systems before hackers—state-sponsored or otherwise—got their hands on similar capabilities.
The revelation landed like a shock. Within days, the Trump administration's entire approach to AI regulation executed a complete reversal. The White House announced agreements with Google, Microsoft, and xAI that would allow the Commerce Department's Centre for AI Standards and Innovation to conduct pre-deployment evaluations of frontier AI models before release. Trump began considering an executive order to formalize government access to these systems. The administration that had spent its first weeks in office tearing down safety frameworks was now scrambling to erect new ones. David Sacks, the AI czar who had so aggressively attacked state-level regulation, was gone from his position. The relationship with Anthropic, previously poisoned, began to thaw.
What made the difference was not ideology but vulnerability. Mythos demonstrated that AI development had reached a threshold where the risks of an unregulated free-for-all had become more consequential than the risks of slowing down development. The tool's capabilities weren't theoretical—they were demonstrable, and they threatened systems that governments and corporations depend on. The shift also reflected a broader recognition, voiced by figures across the industry, that the stakes had fundamentally changed. Elon Musk, in a lawsuit against OpenAI, described the worst-case scenario in stark terms: "a Terminator situation where AI kills us all." Anthropic's Dario Amodei had warned that without careful development, AI "can kill you." OpenAI's Sam Altman had called for regulation of a sector "making the most consequential decisions about the shape of the future."
The timing matters. The US has been aggressively pursuing agentic AI—models that can operate and act autonomously—with the ultimate goal of achieving artificial general intelligence, human-level or beyond. China, constrained by US export restrictions on advanced chips and limited access to venture capital, has taken a different path, focusing on efficiency and integration across its economy. Yet it remains competitive. When Trump meets with Xi Jinping later this month, technology is expected to dominate the agenda. Xi has publicly stated that AI must be "safe, reliable, and controllable" and has been actively promoting global AI safety standards. Trump, by contrast, has already signaled he intends to boast about American AI leadership.
The irony is sharp. China has pursued an open-source model for AI development, leveraging global developers in its competition with the US. The US has pursued a purely commercial approach. In the absence of American leadership on safety standards, that calculus could hand China the most influential role in shaping how the world governs AI development. The Trump administration has finally acknowledged that unsupervised, unchecked AI development poses genuine threats to national economies, critical infrastructure, and financial systems. Whether Trump can actually rein in his wealthiest supporters—the Silicon Valley billionaires who funded his campaign and who see AI as a winner-takes-all race—remains the open question. The sector is expected to attract more than seven hundred billion dollars in investment this year. The answer to that question may determine whether this belated recognition of AI's risks amounts to anything real.
Notable Quotes
The worst-case scenario is a Terminator situation where AI kills us all.— Elon Musk, in court testimony against OpenAI
AI tools have enormous economic value, but if they aren't built carefully, they can kill you.— Dario Amodei, Anthropic
The Hearth Conversation Another angle on the story
So Trump went from calling AI safety "woke" to suddenly wanting government oversight. What actually changed his mind?
Anthropic released a tool called Mythos that could exploit security flaws in every major operating system at machine speed. It could take down power grids, hospitals, financial networks. That's not ideology—that's a tangible threat to infrastructure.
But couldn't he have just ignored it? Why did that particular tool force his hand?
Because it was real and demonstrable, and it made the abstract risk concrete. You can dismiss safety concerns as political correctness. You can't dismiss a tool that could cripple your country's critical systems.
Did Anthropic release Mythos to the public?
No. They gave it to about forty US companies so they could patch their own vulnerabilities before hackers got access to similar capabilities. That restraint actually underscored how serious the threat was.
And now Trump wants to work with Anthropic again, even though he banned them months ago?
Yes. His AI czar is gone, the relationship has thawed, and the administration is considering creating standards-setting bodies for the most powerful models—essentially bringing back guardrails Trump had just torn down.
What happens when Trump meets Xi Jinping next month?
That's the real test. Xi has been pushing for global AI safety standards. Trump wants to boast about American dominance. If they can't find common ground, China might end up shaping how the world governs AI development instead.