The Trump administration, once vocal about slashing red tape around artificial intelligence, now finds itself striking deals with the very companies it once promised to leave alone. Reports point to an imminent voluntary agreement on standards for the most powerful AI systems. This comes after a June executive order that quietly built the framework for government review of frontier models.
Less than six months into the second term, the shift stands out. Early rhetoric from Vice President J.D. Vance championed minimal interference to outpace China. Yet events moved fast. Export controls hit Anthropic in mid-June, freezing access to its latest models for foreign users and forcing a temporary shutdown. OpenAI held back releases of its own. The pressure worked. By late June, the White House lifted those restrictions on Anthropic after the company engaged on security reviews, according to statements from Commerce Secretary Howard Lutnick reported by CNBC and CNN.
Such episodes reveal a pattern. The administration avoids formal regulation. It relies instead on targeted actions and private negotiations. The result looks like de facto oversight without the legislative fight.
At the center sits a June 2 executive order titled “Promoting Advanced Artificial Intelligence Innovation and Security.” President Trump signed it to address cybersecurity risks from advanced models. The order directs agencies including the NSA, CISA, and Treasury to create classified benchmarks for judging when an AI system qualifies as a “covered frontier model.” It also calls for a voluntary framework. Developers can submit models for government evaluation up to 30 days before wider release to trusted partners. Confidentiality protections apply. Nothing in the text creates mandatory licensing or pre-clearance, the order stresses.
“The order asks AI companies to voluntarily submit their most powerful models for the government to test up to 30 days before releasing them to the public,” NPR noted shortly after the signing. Officials framed the move as collaboration, not control.
But the timing matters. The Gizmodo report on July 1, drawing from Financial Times sourcing, described advanced talks for a standards announcement as early as the following week. Those talks involve the Center for AI Standards and Innovation at Commerce and the NSA. Companies in the mix include OpenAI, Anthropic, Google, Microsoft, and Amazon. Meta reportedly held out at first. The standards focus heavily on cybersecurity capabilities. How well models can identify software flaws or enable attacks draws particular scrutiny.
And the benchmarks? They stay classified. Developers receive assessments. The public does not. Shared safety practices across firms may still offer clues about the bar. This opacity draws criticism from those who want transparency in AI governance.
The executive order built on earlier steps. In January 2025, Trump revoked the Biden-era AI rules and ordered an “AI Action Plan.” That plan, released in July 2025, emphasized innovation, infrastructure, and limits on regulation while pushing for objective AI in government use. Later orders targeted perceived bias in federal procurement of large language models. The June 2026 directive marked a pivot toward national security and cyber defense.
Recent reporting adds texture. The Financial Times detailed how the administration accelerated plans for these model standards after interventions in company releases. One source familiar with the discussions told the paper the deal aims to smooth regulatory uncertainty. Without it, firms risk sudden export blocks or withheld approvals.
Industry response splits. Some executives welcome the clarity. Others see it as backdoor pressure. Alex Stamos, a Stanford professor and former security chief, commented on X about Anthropic’s stance after the export lift. The company signaled it had always prioritized safety and now helped shape better communication standards with the government.
But not everyone buys the voluntary label. Legal analysts at firms like Ropes & Gray point out that while the framework disclaims mandatory requirements, the infrastructure of pre-release reviews, trusted-partner tiers, and a cybersecurity clearinghouse could create strong expectations of compliance. “Although voluntary in form, the Order builds significant institutional architecture,” their alert stated, as covered by Ropes & Gray.
Tech Policy Press reported that the order echoes a previously shelved draft. It expands national cybersecurity coordination and increases this “voluntary” oversight. The piece highlighted fears over powerful new systems that can hunt for vulnerabilities at scale. Washington officials grew alarmed. So did allies abroad.
The clearinghouse provision stands out. Within 30 days of the order, Treasury, the National Cyber Director, NSA, and CISA must form a body that works with industry and critical infrastructure operators. It coordinates vulnerability scanning, validation, remediation, and patch distribution. AI models that discover flaws get special attention. The goal is faster defense. The risk is that those same models could weaponize findings.
Frontier developers now face a choice. Engage early with the government on whether their next model crosses the threshold. Grant access under strict nondisclosure terms. Or risk being labeled non-cooperative, with potential consequences for exports, government contracts, or public perception.
This approach contrasts with Biden’s more public risk-management frameworks and voluntary commitments from 2023. Those involved pledges on safety testing and red-teaming. The new setup embeds review inside national security agencies. Classification adds another layer of separation from public debate.
Yet the administration insists on speed. The AI Action Plan from 2025 warned against heavy rules that could slow American leadership. Trump himself has spoken of possible equity stakes in AI firms to let the public share gains, per Financial Times coverage. Innovation and security must advance together.
Critics argue the current path lacks accountability. Without defined public metrics, how does anyone judge if the standards suffice? If Meta or another holdout refuses, what follows? The recent Anthropic episode offers a preview. Temporary controls. Private talks. Then relief after alignment.
So the standards deal, if announced soon, will test this balance. It could set precedents for how the U.S. manages AI capabilities that blur lines between research breakthrough and strategic asset. Companies gain predictability. The government gains eyes on the most dangerous models. Observers wonder whether the arrangement will hold as capabilities surge further.
One thing seems clear. The hands-off era lasted only briefly. National security concerns proved too pressing to ignore. The resulting hybrid of voluntary pacts, classified tests, and selective enforcement now defines the Trump administration’s AI playbook. How far it stretches will shape the technology’s path for years ahead.
WebProNews is an iEntry Publication