Trump Backs AI 'Kill Switch' as Mythos Model Alarms Regulators Worldwide

Anthropic's unreleased AI model, Claude Mythos, has set off a cascade of urgent responses from regulators and government officials across three continents, as financial authorities grapple with the model's unprecedented ability to discover and chain together software vulnerabilities at machine speed. The backlash has now reached the White House, with President Trump publicly backing the idea of government safeguards on artificial intelligence, including a potential "kill switch."

The developments represent the most concentrated regulatory response to a single AI model in history, unfolding across Washington, London, Brussels, and several financial capitals within days of Mythos becoming known to government officials through private briefings.

Washington and Wall Street Scramble | Treasury Summons Bank CEOs

Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened a closed-door meeting with the CEOs of America's largest banks at Treasury headquarters on April 8 to discuss the cyber risks posed by Mythos, according to Bloomberg. Chiefs from Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs were among those summoned, with regulators urging banks to assume a heightened risk posture and accelerate their vulnerability management programmes.

Bessent later described Mythos publicly as "a step function change in abilities, learning capabilities" at a Wall Street Journal event, one of the clearest on-the-record acknowledgements by a senior US official of the model's scale. His characterisation was notably measured given the private tone of the April 8 meeting, which participants described as significantly more alarmed, according to people familiar with the discussions.

The Office of the Comptroller of the Currency has separately issued informal guidance to national banks advising them to review their AI vendor contracts and incident response playbooks in light of what the agency described as "newly identified threat vectors arising from frontier model deployments."

Trump Backs AI Kill Switch | What the President Said

In an interview with Fox Business Network broadcast on Wednesday, President Trump backed the idea of government safeguards on AI, including a potential kill switch. Asked whether AI could affect the banking system's stability, Trump said:

"Yeah, probably. But it could also be the kind of technology that allows greatness in the banking system, makes it better and safer and more secure."

Trump's comment marks a notable shift in tone from the administration's generally permissive approach to AI development. The AI advisory council assembled earlier this year was populated primarily with industry figures who have publicly opposed heavy-handed regulation. Trump backing a kill switch, even in hypothetical terms, signals that the political calculus around AI risk has changed.

White House officials declined to elaborate on whether any formal executive action on AI safeguards is being drafted. The National Security Council has been briefed on Mythos but has not publicly commented on the model.

UK and European Regulators Sound the Alarm

The United Kingdom's AI Safety Institute issued a stark assessment in an open letter to business leaders dated April 15, stating that Mythos is "substantially more capable at cyber offence than any model we have previously assessed." The letter warned that frontier model capabilities are now doubling every four months, up from every eight months previously, a rate the institute described as "outpacing our ability to develop mitigations in parallel."

UK regulators including the Bank of England and the Financial Conduct Authority have begun urgent talks with the National Cyber Security Centre and major financial institutions. Bank of England Governor Andrew Bailey called the capability jump "qualitatively different from prior generations of AI" in remarks to the Treasury Select Committee, indicating that the Bank's existing AI risk framework may need to be rewritten.

In Brussels, the European Banking Authority convened an emergency technical working group on April 14 to assess whether Mythos deployment in EU financial services would require prior regulatory approval under the AI Act's "high risk" category provisions. Sources familiar with the deliberations said the group is leaning toward classifying any deployment of Mythos in critical infrastructure as requiring a conformity assessment before activation, a process that typically takes months.

What Makes Mythos Distinctively Dangerous

The specific capability driving regulatory alarm is Mythos's ability to autonomously discover software vulnerabilities, construct exploit chains linking multiple lower-severity weaknesses into a high-impact attack path, and execute those chains at speeds that outpace any human or automated defensive system currently deployed in enterprise environments.

Traditional vulnerability scanners identify known weaknesses against published databases such as the CVE registry. Mythos, according to briefings provided to regulators, can identify novel zero-day vulnerabilities in live systems, chain them together in real time, and generate working exploit code, a capability previously requiring teams of elite human researchers working over weeks or months.

For financial institutions, the threat model is specific. A Mythos-class attacker, or a criminal group that gains access to such a model, could potentially compromise interbank settlement systems, payment rails, or central clearing counterparties in a timeframe that would not allow a human response. The Bank of England's briefing materials, described by people who have reviewed them, modelled a scenario in which a major clearing bank loses access to its core systems for 72 hours due to a Mythos-assisted attack, causing systemic settlement failure.

Anthropic's Position | Controlled Access and Safety Commitments

Anthropic has defended Mythos, arguing that the model has not been publicly released and is currently accessible only to a small group of vetted institutional partners under strict usage agreements. The company's planned expansion to UK financial institutions is structured as a controlled programme, with Anthropic embedding liaisons directly within each institution's security operations team.

A company spokesperson said: "We develop Mythos specifically to help defenders, not attackers. Every deployment is gated, monitored, and subject to immediate suspension if misuse is detected." The company has also pointed to its internal identity verification requirements as evidence of the seriousness with which it treats access controls.

Critics, including several former national security officials who have been briefed on Mythos, argue that Anthropic's own safety controls are only as strong as the systems protecting the model weights themselves. A theft of the weights, or a nation-state operation that extracts the model through an adversarial prompt attack, would place the capability outside Anthropic's oversight entirely.

What Comes Next | Legislation, Oversight, and the AI Act

The bipartisan Senate AI Working Group, which has been dormant for most of 2026, is expected to reconvene in the coming weeks with Mythos as a central agenda item. Senators from both parties have privately signalled interest in fast-tracking at least a narrowly scoped bill requiring federal agencies to be notified before any AI model with offensive cybersecurity capabilities is deployed in US critical infrastructure.

In the UK, the government has indicated it will accelerate the timeline for its AI Safety Bill, which had been slated for a second reading in June. The bill as currently drafted does not specifically address offensive cyber capabilities, meaning amendments will likely be needed to capture the Mythos threat model.

For Anthropic, the regulatory pressure creates a difficult strategic calculation. The company has built its brand around safety, and the global alarm over Mythos could be read as validation of the risks it has spent years warning about. At the same time, a wave of national legislation requiring pre-deployment approval of frontier models would fundamentally change the economics of the AI industry, and Anthropic would not be exempt.