White House Nears Deal on Voluntary AI Safety Rules
The White House is finalizing voluntary AI safety standards with major labs, with an announcement possible as soon as next week.
Washington and Silicon Valley are about to find out whether a policy built entirely on goodwill actually works. The Financial Times reported on July 2 that the White House is in advanced talks with AI companies to finalize voluntary standards for frontier model releases, with an announcement possible as soon as the week of July 7. Reuters confirmed the same day that Google is among the companies involved, with the talks specifically timed ahead of its planned release of an advanced coding model, Gemini 3.5 Pro, expected sometime this month.
This is not a new law. It is the next step in an executive order signed a month ago, and the fine print of that order is where the real story sits.
A Deal That Could Land Within Days
If an announcement does come next week, it will mark the first real test of a framework the White House has spent a month trying to stand up. The pressure is not coming from Congress. It is coming from a self-imposed deadline the administration wrote into its own executive order, and from a handful of frontier labs racing to ship new models while that deadline closes in.
Google's timing is instructive. A company does not usually schedule policy conversations around a product launch by accident. The fact that Gemini 3.5 Pro's release window overlaps with these talks suggests Google wants clarity on the rules before it ships, not after regulators start asking questions about a model already in the wild.
What the June Executive Order Actually Requires
On June 2, President Trump signed "Promoting Advanced Artificial Intelligence Innovation and Security," an order that does two main things. First, it directs federal agencies to harden cyber defenses across government systems and critical infrastructure, with binding deadlines attached. Second, and more relevant here, it creates a voluntary framework under which developers of what the order calls "covered frontier models" can give the federal government up to 30 days of pre-release access to a model before it reaches other trusted partners.
The National Security Agency, working with a group of other agencies, is tasked with building a classified benchmarking process to decide which models actually meet that "covered frontier" threshold. Developers will not see the criteria. The order is explicit that none of this creates a mandatory licensing or preclearance requirement, a distinction the administration has repeated at every stage of drafting. Companies can say no. The question the coming weeks will answer is how many actually do.
Why 30 Days, Not 90
Earlier drafts of the order had set the government access window at 90 days, according to legal analysis from A&O Shearman. The final version cut that to 30, a change described as the most significant revision between draft and signature. That is not a minor tweak. Three months of federal access to an unreleased model is a very different commercial risk than one month, and the shorter window reads as a direct concession to labs worried about competitors, including Chinese developers, shipping while American models sit under review.
That compromise tells you exactly which political faction won the internal argument. National security officials wanted a longer look. The administration's anti-regulation wing won a shorter one. Thirty days is what happens when neither side gets to walk away with everything it asked for.
The Incident That Set This in Motion
None of this policy exists in a vacuum. The trigger, widely reported at the time, was Anthropic's Mythos Preview model demonstrating an ability to find critical vulnerabilities in widely used operating systems that had gone undiscovered by human researchers for years. NBC News described the reaction inside Washington as sending shockwaves through the national security establishment. That single demonstration reframed frontier AI, in the eyes of policymakers, from a productivity story into a cybersecurity story almost overnight.
That reframing explains why this executive order lives inside the national security apparatus, run by the NSA and CISA, rather than inside a consumer protection or antitrust framework. The government is not worried about AI taking jobs or spreading misinformation here. It is worried about AI finding the next unpatched hole in the software running hospitals, banks, and utilities before anyone else does, and it built a policy response around that one fear.
The Loophole Nobody Has Tested Yet
Here is the structural problem this framework has not solved: it asks for voluntary cooperation with no penalty for declining. A lab racing to beat a competitor to market has every incentive to skip a 30-day government review it can legally ignore, and nothing in the order changes that calculation. Confidentiality and intellectual-property protections are promised on paper, but handing an unreleased model to federal reviewers still means exposing a company's most valuable asset to people outside the company, penalty-free refusal or not.
Whether labs actually say yes when asked is the one variable this entire order leaves in private hands. An agreement announced next week would answer that question for the current round of releases, but it would not settle it permanently, because every future model launch resets the same choice for whichever lab ships it.
What to Watch as August 1 Approaches
The order's own deadline lands on August 1, when the voluntary framework's mechanics are supposed to be fully designed. If the White House and AI companies do reach an agreement next week, it will likely cover only the broad shape of participation, not the granular detail of how "trusted partners" get selected or how classified benchmark thresholds get set, both of which remain undefined in the order's public text.
Google's involvement ahead of a specific model launch is the clearest signal yet that at least one major lab sees value in getting ahead of the framework rather than waiting to be designated a "covered frontier model" after the fact. Whether that logic holds for labs with less patience for a 30-day pause, particularly any racing a rival to market, is the part of this story that has not happened yet.
Written by
Mr. Aayush Bhatt
Software Engineer interested in how models work and where they fail.