White House considers reshaping U.S. AI security approach ahead of China visit
Trump administration appears poised to reshape the U.S. approach to AI security ahead of President Trump's trip to China, considering government vetting of models and AI safety talks with Beijing.
Objective Facts
The Trump administration appears poised to reshape the U.S. approach to AI security ahead of President Trump's trip to China next week, which could be the turning point for how the Trump White House handles the proliferation of the most advanced AI models in the world. The administration is studying possibly an executive order to give a clear roadmap for how future AI that potentially create vulnerabilities should go through a process so that they're released in the wild after they've been proven safe, just like an FDA drug, according to National Economic Council Director Kevin Hassett. The shift was triggered by Anthropic's Mythos model, which found thousands of vulnerabilities in operating systems and web browsers, raising concerns that cybercriminals or hostile foreign agents could penetrate computer systems and compromise basic computer code underlying public safety, national economies and military security. At the Trump-Xi summit in Beijing, the United States and China agreed Thursday to establish a protocol on best practices for artificial intelligence safety, with Treasury Secretary Scott Bessent stating they would set up a protocol in terms of how to go forward with best practices for AI to make sure non-state actors don't get a hold of these models. This represents a major reversal, as accelerationists previously in command of Trump's AI policy were firmly opposed to any discussions with China and used the "China card" as a reason for opposing domestic AI regulation.
Left-Leaning Perspective
Left-leaning and centrist outlets have cautiously welcomed the Trump administration's shift toward AI oversight but argue the approach remains inadequate without truly independent government testing. The New York Times published commentary from Dean Ball and Ben Buchanan supporting "appropriate guardrails on AI development," beginning with mandatory audits of developers' safety claims "by independent expert bodies overseen by the government," and calling for "tighter controls on the critical technologies that China needs," like AI chips. However, Bloomberg columnist Parmy Olson, cited in The Week, criticized that "the U.S. under Trump will likely never lead the way on regulating AI," with Trump's proposed working group including tech executives who could essentially "write the rules meant to police them." TechPolicy.Press analyst Emma Hatheway argued that while Anthropic's decision to withhold Mythos appears responsible, a safety regime that depends on CEO discretion is insufficient, and a federal review directly influenced by industry tech giants would not be better. Brookings Institution expert Darrell West, quoted by The Register, suggested Trump is essentially returning to Biden's policy, while the concern persists that meaningful oversight cannot emerge from a working group co-designed with the companies being reviewed. Left-leaning coverage emphasizes that independent evaluation bodies like the London-based AI Security Institute represent "the best-funded AI vetting agency in the world," as noted by The Week, and that only such independent entities earned Anthropic's trust with Mythos. Northeastern University's Alan Mislove told the university's hub that he is "glad that the Trump administration is realizing the significant risks of AI," though this remains far short of the comprehensive regulatory framework many progressives advocate. The broader left perspective suggests the shift is welcome but insufficient without Congress passing binding legislation and without meaningful separation between government evaluators and the companies they oversee. Left-leaning outlets generally omit or downplay the economic competitiveness argument for rapid AI development and the risk that overseas regulation could disadvantage American firms. They focus primarily on safety and equity concerns rather than on whether vetting could slow U.S. innovation relative to China.
Right-Leaning Perspective
Right-leaning and libertarian policy analysts are sharply critical of the proposed AI vetting regime, framing it as a dangerous reversal of Trump's pro-innovation stance that will harm American competitiveness and favor entrenched players. Jennifer Huddleston and Juan Londoño at the libertarian Cato Institute, cited in The Hill, warned that requiring pre-launch approval "was criticized as heavy-handed and anticompetitive when included in the Biden administration's executive order on AI." The American Enterprise Institute published a detailed critique arguing that "a mandatory government AI vetting regime would likely do little to enhance security, while creating significant harm to innovation and competition," with compliance costs creating barriers to entry for startups compared to established firms like Anthropic and OpenAI. The Daily Signal, a conservative outlet, noted that "there are differences of opinion within the administration about how strong the vetting process of new models should be," with some officials preferring a light touch. The Register's coverage framed the policy shift sarcastically: "Trump jumps from 'anything goes' to 'strict regulation' AI policy," quoting Trump's earlier statement about letting the AI "baby thrive" rather than "stop it with foolish rules and even stupid rules." Right-leaning outlets generally praised the original Trump position that favored innovation, with Vice President Vance's earlier statement in the "All-In" podcast that "America is done with the overregulating" being cited as the prior administration position. Right-leaning analysis emphasizes that licensing schemes and pre-approval reduce innovation by standardizing products to meet regulatory requirements and that the costs of compliance favor incumbents. The AEI analysis argued that "responsible AI development" works best through company self-policing backed by litigation accountability, allowing "industry standards to emerge, shaped by practitioners rather than regulators, and refined by experience rather than conjecture." This perspective suggests that bottom-up, market-driven approaches adapt better to rapidly changing technology than government regulations. Coverage notes that even current voluntary testing agreements with companies like Google, Microsoft, and xAI have already occurred without mandatory vetting, suggesting the existing approach is sufficient. Right-leaning outlets generally understate or omit security vulnerabilities demonstrated by Mythos and downplay the national security rationale offered by administration officials like Kevin Hassett and Scott Bessent for the shift.
Deep Dive
The Trump administration's reshaping of U.S. AI security policy represents a genuine pivot driven by two converging factors: Anthropic's Mythos model demonstrations in April 2026 proved that frontier AI can rapidly identify and exploit critical cybersecurity vulnerabilities, and the appointment of Treasury Secretary Scott Bessent and National Economic Council Director Kevin Hassett after David Sacks's March departure created space for national security officials (the NSA, intelligence agencies, and the White House National Cyber Director) to influence AI policy. The administration initially entered office with an explicitly pro-innovation, anti-regulation posture, with Vice President Vance declaring in the "All-In" podcast that America was "done with the overregulating" and Sacks criticizing AI "doomers." However, when Anthropic voluntarily limited Mythos access to only about 50 critical infrastructure partners, and disclosed that the model could find thousands of zero-day vulnerabilities in major software, the White House pivot became inevitable. The convergence of Mythos revelations with Trump's planned May 14 summit with Chinese President Xi Jinping created additional urgency: the administration needed a policy position on both domestic vetting and international AI governance before the meeting. The left correctly identifies that the proposed vetting approach—a working group including tech executives and government officials—remains weaker than independent oversight models (like the UK's AI Security Institute, which Anthropic actually trusted with Mythos). Their criticism that companies could influence rules meant to police them has analytical weight: compliance costs for pre-release vetting do disproportionately burden startups, and there is historical precedent of regulated industries capturing regulatory processes. However, the left understates that Mythos posed genuine, non-partisan national security concerns and that some form of pre-release assessment has already begun voluntarily—Google DeepMind, Microsoft, and xAI agreed to pre-deployment testing with CAISI within days of the executive order discussions. The right correctly notes that mandatory licensing schemes can reduce competition, and there is real economic cost to compliance that may favor incumbents. However, the right minimizes the actual threat Mythos represents and suggests that self-regulation and litigation accountability suffice, when the vulnerability discovery capability Mythos demonstrated is qualitatively different from previous AI risks. The right also underestimates the bipartisan public demand for vetting: polling shows 47% of Republicans strongly support mandatory testing and over 80% of Americans overall favor it, suggesting the policy shift reflects public pressure, not just technocrat whim. On the China dimension, a crucial strategic question remains unresolved: whether the U.S.-China AI safety protocol discussion represents a genuine risk-reduction dialogue or whether China will leverage it primarily to gain access to U.S. technology insights while maintaining its own AI race unencumbered. Council on Foreign Relations analysts note Beijing's historical pattern of extracting technology access through arms-control dialogue while maintaining strategic flexibility. The fact that Bessent framed U.S. willingness to talk in terms of "we are able to have wholesome discussions with the Chinese on AI is because we are in the lead" suggests the administration recognizes it can only negotiate from strength—but also that the window for such negotiations may close as China narrows the AI performance gap (which Stanford researchers say has "effectively closed"). Neither left nor right fully confronts the risk that any formal AI safety agreement might actually codify Chinese AI development levels at current or near-current capability, inadvertently preventing the U.S. from deploying more advanced systems later. What to watch next is whether Trump and Xi actually establish a working dialogue (with separate government agencies, evaluation criteria, and enforcement mechanisms still undefined), whether the domestic vetting process blocks any major model releases in practice (the real test, since reviewing without consequential blocking power is ceremonial), and whether Congress moves from bipartisan polling support for vetting into actual legislation that would make this binding rather than executive-order vulnerable.
Regional Perspective
At the Trump-Xi summit in Beijing, the United States and China agreed Thursday to establish a protocol on best practices for artificial intelligence safety, with Treasury Secretary Scott Bessent disclosing the agreement and stating they would set up a protocol in terms of how to go forward with best practices for AI to make sure non-state actors don't get a hold of advanced models. However, how the proposed AI safety protocol will be structured, governed, or enforced remains unclear, and both governments have yet to release formal documentation of the agreement. So far, the only meaningful public comments on the topic have come from Treasury Secretary Scott Bessent, who indicated that future bilateral discussions on AI safety and guardrails may focus initially on preventing the most advanced AI models from falling into the hands of nefarious non-state actors, with crucial questions remaining about when talks begin, who leads them, who is included, whether they remain confined to official Track I diplomacy or expand into Track 1.5 and Track II dialogues, and whether the agenda eventually broadens beyond narrow AI safety concerns. According to Council on Foreign Relations analysis by Chris McGuire, who led U.S.-China AI policy under Biden, Beijing will not negotiate in good faith on AI safety, as the Chinese government's willingness to make and abide by robust international commitments on AI safety is low, viewing such dialogues as opportunities to increase its access to technology that China needs to catch up to the United States in AI. The Chinese government's perspective on AI safety cooperation is consistent with its longstanding refusal to agree to substantive arms control measures with the United States, with China viewing arms control with extreme skepticism and having a poor track record of abiding by arms control commitments it does make. In the U.S.-China tech race, Beijing has underscored AI control from the start while the U.S. seems only now to be taking it seriously. International perspectives diverge sharply on whether this represents genuine cooperation or strategic positioning. According to Stanford researchers, the U.S.-China AI model performance gap has effectively closed, shifting the urgency for both nations to establish guardrails. The London-based AI Security Institute represents the best-funded vetting model globally and is the only government agency Anthropic trusted with Mythos access, suggesting that international oversight mechanisms exist but were not extended from Washington. The fact that Beijing agreed to AI safety talks signals both countries recognize shared interest in preventing non-state actor access to dangerous models, though the Biden administration's prior dialogue with China on shared risk saw Beijing indicate they weren't really interested in negotiating in good faith on shared risks, with their number-one priority being catching up to the United States in AI.