White House Finalizes Secret AI Safety Framework, Shares Only with Select Tech Companies

The White House Is Keeping Its AI Cybersecurity Framework Secret

The Trump administration has finalized a plan to address the cybersecurity risks posed by increasingly capable artificial intelligence models, a White House official confirmed to WIRED. But at least for now, it's deliberately keeping the details under wraps, people familiar with the matter tell WIRED. The Trump Administration invited staffers from OpenAI, Anthropic, Google, Meta, Nvidia, and other leading AI companies to the White House on Tuesday to share an overview of its new AI oversight framework, the people said. AI developers will have the ability to voluntarily submit new models to the federal government up to 30 days ahead of their public release. The White House will then vet their cyber capabilities according to a classified benchmarking system and share the AI models with federal agencies and trusted corporate partners. The White House isn't sharing more information about its testing criteria or which AI models will be covered by the framework, though open models will reportedly be excluded, according to Axios. That has left smaller AI startups, safety advocates, and third-party researchers in the dark about crucial aspects of how the federal government is addressing the cyber risks posed by advanced AI systems. Some argue the secretive process will give an advantage to larger companies. "They're essentially creating an entrenchment program for the big AI model providers, which are now considered the most frontier," says a person familiar with the White House's discussions with AI labs, who requested anonymity to discuss confidential matters. "This creates an economic incentive program for critical infrastructure just to use them, and leaves out smaller startups." The White House did not respond to requests for comment. The Trump administration may be keeping its AI security framework confidential because of national security concerns. A second White House official, who requested anonymity because they were not authorized to speak to the media, emphasized that the new framework is intentionally narrow and focused exclusively on the cybersecurity capabilities of the most advanced models on the market, such as Anthropic's Fable and OpenAI's ChatGPT 5.6. But some AI safety advocates tell WIRED that any rules AI companies are being held to should be made public to ensure third-party groups can keep them accountable. "This is far too important an issue to be hidden behind a cloak of secrecy," says Brad Carson, president of the nonprofit Americans for Responsible Innovation and cofounder of the pro-regulation Public First Action super PAC, which has funding from Anthropic. "This is not a handshake deal with tech companies. It's the rulebook for ensuring they don't endanger the public. If only tech companies know what's in the rulebook, it doesn't work." Cyber Concerns The oversight framework stemmed from an executive order President Donald Trump signed earlier this year designed to address the cybersecurity risks of new AI models. In recent months, Trump officials have grown increasingly alarmed about the hacking capabilities of cutting-edge AI systems, which they worry could pose a serious risk to national security. Those fears escalated over the last two weeks when OpenAI and Anthropic said they discovered their AI models had unknowingly bypassed controls and hacked into third-party services during internal testing. The House Committee on Homeland Security sent a letter to OpenAI CEO Sam Altman last week requesting that he brief lawmakers about how one of the company's AI agents breached the platform Hugging Face. "This incident really is a wake up call for people that agent capabilities have now reached this level," Dawn Song, vice president of AI research at Meta, said during a panel discussion on Saturday at the University of California, Berkeley, where she is also a professor, referring to the Hugging Face breach. The Trump Administration's new framework is an attempt to strike a balance between promoting competition in the AI industry and maintaining safety. The executive order notes that it should not be seen as a "mandatory licensing regime," but critics have argued that the Trump Administration's opaque process has created just that. "The regulations necessary to prevent the catastrophic risks presented by uncontrolled AI and superintelligence should not be voluntary," says Conor Leahy, executive director of ControlAI, a non-profit focused on countering AI risks. "This action admits the danger, but leaves the burden of safety in the hands of companies that have an incentive to proceed at full speed with disregard for the wellbeing of the public." Weighty Matters For the last year and a half, White House officials have been wrestling with how to mitigate the risks of advanced AI without stifling American innovation or ceding ground to China. President Trump returned to office promising to take a hands-off approach to AI, but his administration has shown a growing willingness to intervene on the issue. In June, for example, it took the unprecedented step of placing temporary export controls on Anthropic's most advanced AI models over cybersecurity concerns. The decision prompted Anthropic to take its models offline altogether until it could reach an agreement with the Trump administration. Later that month, OpenAI said it was delaying the rollout of its latest AI model, GPT-5.6, in response to a request from the White House. The saga prompted outcry from tech executives in Silicon Valley, who worried that excessive regulation could lock in a handful of companies as the winners of the AI race. A key issue US officials have debated is whether to restrict the distribution of open-weight AI models, which can be freely downloaded and modified. A number of the leading ones are developed by Chinese companies and have become popular among researchers and startups. Some voices in Washington have called for a ban on Chinese open-weight models while others have advocated for promoting US open models as an alternative. More than 80 companies signed an open letter last week organized by Nvidia that asked the US government to defend open-weight AI models. On Tuesday, Nvidia and the same coalition of companies launched a new project called SAFE, or Shared AI Findings Exchange. The goal is for tech companies to "confidentially collect and analyze AI incidents and near misses, identify recurring control failures and publish evidence-based operating recommendations that reduce systemic risk," according to a blog post Nvidia published. In addition to Nvidia, Hugging Face and Red Hat have agreed to participate in the project, and The Linux Foundation called on other organizations to make their own open-source contributions to it. "As an industry, we want to have this conversation out in the public," Justin Boitana, vice president of enterprise AI at Nvidia, said in an interview with WIRED. The goal is for SAFE to be "governed independently, with no single company or industry segment controlling its findings." Boitano declined to say whether Nvidia has discussed the White House's new framework with Trump officials. However, Boitano says, "I think [our] framework is one to look at," referring to SAFE. During the Agentic AI Summit at Berkeley over the weekend, OpenAI cofounder Wojciech Zaremba, who currently serves as the head of AI resilience at the company's philanthropic arm, said that the AI industry is "entering a new era." "Imagine what would happen if, all of a sudden, the locks to your house stopped working," Zaremba said during the same panel discussion where Meta's Dawn Song spoke. "That's the era that we are entering with cybersecurity... My guess is that it will be chaotic."

The White House's New AI Safety Framework Is None of Your Business

The Trump administration has cooked up a framework meant to ensure that new and powerful AI models don't pose a threat before being released to the public. According to Axios, the public will not be allowed to see it. The White House plans to keep its rubric for evaluating AI models confidential, sharing it only with the companies whose models will be judged against it. With incident after incident of AI models breaking containment and allegedly hacking into systems (which has somehow turned into a marketing tool for these companies who want to brag about how smart and capable their rogue technology is), it'd seem like a pretty ideal time to not only establish a set of standards, but to share them so that everyone knows what to expect and where the bar has been set for safety. The Trump administration is apparently cool with sticking with one out of two -- and even the one is pretty opaque. Per Axios, the White House finalized its AI oversight framework yesterday and had some industry players in the building on Tuesday to give it a look. Beyond the small group granted access, no one really knows what's in it. Per a Politico report, the few who have taken a peek work at Anthropic, Google, Meta and OpenAI. The only real indicator we have as to what the framework entails comes from the initial executive order, signed by Trump in June after some industry pushback initially caused him to get cold feet. That order requires the government to "develop and maintain a classified benchmarking process to assess the advanced cyber capabilities of AI models." It also calls for those benchmarks to be made available to developers and researchers "as appropriate," which the White House is apparently interpreting as "whenever we feel like it." Of course, it kinda doesn't matter what's in there, considering the whole thing is a voluntary process, anyway. So far, AI companies have chosen to comply, likely as part of their ongoing "We swear we're responsible" posturing that they perform to counterbalance their "Oops, our model did a cybercrime" campaign. And they may continue doing so as long as they do not find the secretive testing process particularly burdensome. But if that changes, companies could simply stop submitting their models for evaluation. There's not really a lever in place should that happen, unless they're hiding that, too.

The White House's plan to vet potentially dangerous AI is cloaked in secrecy

A Trump administration framework on AI testing leaves a lack of transparency - and plenty of open questions After months of talking with tech industry leaders, the Trump administration finalized a framework this week for how it will test new artificial intelligence models for safety and cybersecurity risks. So far, the White House is keeping details of the framework private, in a blow to transparency and potential boon for secretive AI companies. On Tuesday, staff from OpenAI, Anthropic, Meta, Google, Nvidia and Microsoft attended a private meeting with White House officials to review the AI framework. Multiple outlets have since reported that although the volunteer vetting process for new AI models has been settled, the White House does not plan to release its policy publicly and will only share testing criteria with a select few tech companies. The Trump administration and tech industry's opaque process for how they will check if new AI models threaten global cybersecurity has left numerous questions about how the framework will operate. It remains unclear what level of scrutiny models will face and what safety benchmarks they must meet, leaving businesses, foreign governments and the public in the dark. The White House, OpenAI and Anthropic did not respond to a request for comment. Discussions over creating an AI cybersecurity framework began earlier this year after the release of Anthropic's Mythos model, which the company withheld from public release in April over concerns that it could be used to hack into IT and financial systems. The model's capabilities sparked a small geopolitical crisis over cybersecurity, as well as spurred the Trump administration to reconsider its hands-off approach to AI regulation in favor of slightly more oversight. In June, the White House issued an executive order that called for AI companies to voluntarily submit their new models for government review up to 30 days before release. The order was a watered-down version of initial notions to make some parts of the vetting process mandatory. Tech moguls including Elon Musk and Mark Zuckerberg reportedly personally lobbied Trump against the mandate. The order also left an August deadline for determining the framework, which the White House and AI companies now appear intent on keeping to themselves. The murky nature of the framework increases uncertainty for businesses reliant on AI models, as well as on foreign governments increasingly worried that frontier AI models can pose unexpected security risks. It also means that outside researchers and cybersecurity experts have little visibility into how the government is assessing new AI models, with only companies such as OpenAI and Anthropic privy to the process. It is also unclear which companies' models will be subject to government review, since the executive order doesn't define what qualifies as the advanced AI it targets. Open source models, which are free to download and use, for instance, will be excluded from the framework, according to Axios. The framework comes amid persistent security concerns surrounding new AI models. Over the past month OpenAI, Anthropic and Meta have disclosed that their new models hacked into outside organizations during what were intended to be isolated security tests. OpenAI and Anthropic also agreed to delay the release of new AI products over cybersecurity concerns and fear that their models could be used to hack into financial systems or otherwise create harm. Earlier this year, the Trump administration ordered the Center for AI Standards and Innovation to stop issuing public reports on AI model assessments while it worked out a framework. It's unclear whether those reports will resume now that the framework is finalized.

Scoop: Inside Trump's AI framework

Why it matters: The voluntary framework will determine how the Trump administration reviews advanced AI models before release, but the White House isn't making it public. What's inside: The AI framework reviewed on Tuesday defines a covered frontier model as closed-source with state-of-the-art capabilities and national security risks, multiple sources briefed on meetings held at the White House said. * There is no clear definition of what is considered state-of-the art or a national security risk. * Open models are excluded, and the framework explicitly says nothing in it should be interpreted as restricting open models once they've been released. Zoom in: During the 30-day pre-release government review period, employees would be limited from accessing models. * The models would be stored in high-security environments and there would need to be detailed logs on who is accessing the models. * The review process will include various administration officials rather than a single office or agency. The big picture: The White House does not plan to publicly release the framework, Axios previously reported. * The White House on Tuesday held staff-level meetings with industry to review the framework, and companies that weren't invited remain in the dark about its contents. * It's also unclear which "trusted partners" will get early access to advanced models under the framework, including whether any foreign governments would qualify. Context: The executive order explicitly says the benchmarking process to assess advanced cyber capabilities of AI models will be classified. * There is no requirement in the order to publicly release the voluntary framework. What's next: Companies on Tuesday were reportedly encouraged to share models with the government that are as close to public release as possible instead of early stage ones, according to people familiar with the discussions.

White House won't publicly release AI model evaluation framework it reviewed today with Meta, Nvidia, Microsoft, OpenAI, Anthropic, variety of smaller companies | Fortune

The White House has no plans to publicly reveal the framework it's been working on for how it will vet frontier AI models prior to release. Instead, the details will be kept under wraps, only known to a select group of companies that may choose to participate in the process, which is voluntary. Several major tech companies traveled to Washington, D.C. today for a meeting to review the current draft of the proposal. Attendees included Meta, Nvidia, Microsoft, OpenAI, Anthropic, and a variety of smaller companies, according to sources familiar with the matter. Fortune is first to report that Microsoft was in attendance. The administration issued an executive order on June 2 mandating the creation of this framework within 60 days, or by August 1. The directive seeks to define which models are eligible for review, and instructs the AI labs that they have "up to 30 days" to submit them to the government prior to their public release. The secrecy surrounding the framework may not instill public confidence in the government's ability to vet and secure powerful AI models, especially after OpenAI confirmed its models hacked into another company, Hugging Face, last month. Anthropic later confirmed its models had done the same three times. The fact that the process is voluntary raises questions about how the administration will enforce it. Per the executive order, the framework is not "mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models." Chris McGuire, Senior Fellow for China and Emerging Technologies for the Council on Foreign Relations, called the decision to keep the framework behind closed doors "baffling." "We can't have secret, voluntary rules to regulate the most important tech in the world," McGuire wrote on X. It's unclear if the administration is operating behind closed doors for national security reasons, because it does not want input from outside researchers and experts, or for some other reason. The U.S. government has already been working with major AI companies to review their latest model releases. In June, it effectively took Anthropic's Mythos 5 and Fable 5 models off the market, subjecting them to export controls, and then worked with the company to fortify security before making them available. Then, the government worked closely with OpenAI ahead of its July 9 debut of GPT-5.6. On July 21, Google said it had made its 3.5 Flash Cyber model available to the government ahead of release as well. Current discussions on Capitol Hill likely aim to formalize these engagements. It's unclear if the framework is finalized or still in progress. In the meeting today, attendees floated the idea of a future event related to the proposal, perhaps to continue discussing it.

White House plans to keep AI framework under wraps

* Keeping it private means companies, policymakers, researchers and U.S. allies outside the process will be left guessing how the administration plans to implement one of its key AI policies. Driving the news: The White House on Tuesday held staff-level meetings with industry to go over the recently completed framework. Companies that weren't invited remain in the dark about its contents. * The AI framework comes as industry grapples with high-profile cyber-attacks and rapid advances by Chinese AI developers, reigniting a debate on how to deal with open-source models. * It's also unclear which "trusted partners" will get early access to advanced models under the framework, including whether any foreign governments would qualify. The European Union declined to comment and the U.K. did not respond to multiple requests for comment. * The White House declined to comment. Zoom in: Open-source models were discussed, sources familiar said, without elaborating. Nvidia staff participated in the meetings, the sources added. * Nvidia CEO and open-source advocate Jensen Huang was in D.C. last week meeting with Commerce Secretary Howard Lutnick and President Trump. Catch up quick: The voluntary framework, outlined in a June executive order, is intended to give AI developers a process for working with the government to determine whether models under development fall within its scope.

White House, AI Firms Keep Safety Framework Talks Private

The White House met with representatives from leading artificial intelligence companies today to discuss a safety framework for the government to review frontier models prior to launch, although there's no plan as yet to make the details public fare. The Trump administration teamed up with executives from a number of firms that included Anthropic PBC, OpenAI PBC, Google LLC, Meta Platforms Inc. and Nvidia Corp. The conversation on safety follows a number of incidents in which advanced AI models seemingly went rogue and have been adjudged to have become a major safety concern as companies race to build the most advanced systems. Last week, U.S lawmakers called for the introduction of a government-held "kill-switch" of sorts after a model Open AI was testing discombobulated the firm by breaking out of its testing environment and hacking the open-source AI platform Hugging Face Inc. Just a few days later, Anthropic reported a similar event in which two of models breached the sandbox and went on a hacking spree. Today's meeting was supposed to come to a decision that will mean the government has more control over AI development in an effort to prevent something being released that could cause significant damage. Companies will be expected to give the government access to models at least 30 days before they hit the public sphere, although this will be voluntary and reportedly only for models with the most advanced hacking capabilities. "The Administration's expected action this week on frontier AI could be an important step toward closing the gap between innovation and governance: a clear, credible, national framework for evaluating the most advanced AI systems, with defined criteria, timelines, and a process that allows them to be deployed safely and quickly," said Chris Lehane, Open AI's Chief Global Affairs Officer, in a blog post. With the process being voluntary, it's difficult to know exactly how it will work. Chris McGuire, Senior Fellow for China and Emerging Technologies for the Council on Foreign Relations, called the secrecy around the framework "baffling" on X. "We can't have secret, voluntary rules to regulate the most important tech in the world," he added.

Trump administration's closed-door AI framework catches tech policy sector off guard

The Trump administration's decision to keep its new framework on government AI testing out of public view has blindsided the tech policy community, who waited months to see the controversial process ironed out. The framework was expected to detail the process and benchmarks for government testing of private AI models ahead of their public release. In turn, industry analysts hoped it could clear up the months of confusion over the administration's influence in private model releases. But when the White House hosted a handful of the largest frontier AI labs to brief them on the framework, attendees were told not to expect it to be released to the public, according to multiple sources familiar with the meeting. "We were hoping for clarity around what and some certainty around how the process would work, and the fact that the framework is not going to be public makes that harder to get right," Neil Chilson, the head of AI policy at the Abundance Institute, said. Representatives from labs Anthropic, Meta, OpenAI and Microsoft attended last week's briefing, according to the sources familiar, while Forbes reported staff from Nvidia and other small companies attended. It is not clear whether non-attending firms will be able to access the framework. The framework was a required part of President Trump's executive order in June, which officially launched a new process for AI companies to voluntarily share their models with the government for up to 30 days before releasing them publicly. The order gave multiple agencies 60 days to develop a classified benchmarking process to evaluate the cyber capabilities of AI models. While benchmarks were expected to stay behind closed doors, the order also called for agencies to develop a "voluntary framework," laying out how a company can be designated a "covered frontier model," and collaborate with the federal government to select "trusted partners" that also have early access to new models. Axios reported the framework defines a covered frontier model as closed-source with state-of-the-art capabilities and national security risks but the capability levels or national security risks were not specified. The Foundation for American Innovation (FAI), a center-right think tank focused on technology and policy, filed a Freedom of Information Act request to the White House's Office of Cyber Director for the release of the framework, according to Samuel Hammond, FAI's chief economist and AI policy director. Hammond said he is concerned the framework could be "thin" based on this report. "I worry that it lets the White House check a box that they're doing something on AI in a way that neglects some of the more extreme issues that are coming down the pipe," Hammond told The Hill. "The government can classify certain things if they need to," Hammond added, "But final agency decisions like this are textbook public records." Chilson, who served as the acting chief of technology at the Federal Trade Commission during Trump's first term, echoed these concerns, also warning it could "really slow roll" the release of American models while the government gets to keep its access. "That type of uncertainty and that type of ability to choose who gets to have powerful models and who doesn't, I just think that's not appropriate to be done behind closed doors by a single branch of the government," Chilson said. The framework follows a pivotal 60-day period for the industry, filled with new model delays and "rogue" AI models that upped anticipation for the official rules of the road. Within weeks of Trump signing the executive order, both OpenAI and Anthropic dealt with delayed model rollouts amid the government's cybersecurity concerns. Anthropic was hit with an export control on its newest Fable and Mythos models in June, forcing the company to temporarily take the models offline while additional testing occurred. While OpenAI did not face export controls, it announced weeks later it would delay the public rollout of GPT-5.6's Sol, Terra and Luna at the request of the Trump administration. Both firms eventually released their models following testing, but nearly two weeks later, OpenAI agents' hacking of another company sent new shockwaves through the industry. OpenAI revealed late last month its Sol model, and another unreleased model, went "rogue" during cybersecurity testing and escaped an isolated sandbox to access the internet and eventually breach the systems of Hugging Face, a technology startup. Anthropic and Meta have since disclosed other incidents in which a misconfiguration by the cybersecurity company Irregular allowed models in isolated environments to access the internet and hack other systems. These disclosures have stoked fears around the growing capabilities of AI when on its own, or used by a bad actor. "There's just a lot of public mistrust or distrust of the AI companies and concern about AI per se," Hammond said. "I think there are benefits to just this being more public for creating greater public trust in whatever process is being implemented rather than it be sort of a shadowy, closed-door thing that only a handful of companies are privy to at the expense of both the public and their competitors." Trust and sentiment about AI has drastically fallen among Americans over the past year amid broader concerns about job losses, misuse and data centers. A Bentley University-Gallup survey released last month showed 39 percent of respondents said they believe AI does more harm than good, compared with 31 percent who said so in 2023. "Reports that the US government won't publish its AI regulatory framework are baffling," Chris McGuire, who led U.S.-China AI policy at the National Security Council in the Biden administration, wrote social platform on X. "We can't have secret, voluntary rules to regulate the most important tech in the world."

Trump Administration Plans to Review Top Closed AI Models for Security Risks

What's interesting is that the White House does not plan to review open-source AI models, which makes us wonder whether Chinese models are being excused White house officials held meetings with representatives of top AI companies earlier this week and reportedly informed them that the federal government would review only some top closed frontier models for potential security risks. What this means is that open-weight models like those coming from China may not fall under its review ambit. A report published by New York Times quoting sources said that the review essentially would involve only closed models that do not publish their underlying code. Currently, we have OpenAI, Anthropic and Gemini on this list. While open source models may not be part of the current list, things could change as the technology advances. Donald Trump, which had promised a free and open path to AI development during the run-up to his second term at the White House, has hardened the stance in recent times, even blocking export of Anthropic's latest Claude Mythos model and its OpenAI equivalent GPT-5.6. Fable 5, which was put under guardrails by Anthropic also met the same fate. What possibly got the White House to step in with a firm hand was reports of autonomous cyberattacks by a rogue OpenAI agent on Hugging Face and a few other companies. And when Anthropic claimed that their agent too had broken into some hapless secure database, which they only realised after the GPT-5.6 incident, the federal authorities just speeded things up. The latest move, kept under wraps by the White House, appears to suggest that the hands-on approach towards regulating AI is here to stay with clear oversight getting established on AI labs like Anthropic and OpenAI. What caused them to delay regulatory decisions on the technology and potential threats from Chinese companies is yet unclear. An obvious reason could be the recent concern over blocking open-source models among the tech industry led by the likes of Nvidia, Microsoft and others. Their Open Secure AI Alliance (OSAA) is pushing for a change in narrative around regulation of open-source models. With over 120 signatories to an open letter, the group is now working on new policy proposals. These companies have taken pains to point out that open source models could become a boon for enterprises trying to get a better return-on-AI-investment as they can download the code underneath the models and modify them. Moreover, these companies can use their data on top of these models without having to share it with the AI models, something that companies like Anthropic and OpenAI force users to do. It remains to be seen whether the Trump administration's efforts to draw up an AI framework would shape the future of innovation in this ecosystem. Till date, companies merely launched new AI models at their convenience, though Anthropic overturned this process by keeping its Mythos 5 under wraps over security fears and its prowess at hacking to fix really old codes. When the Trump regime moves in with its formula, one can expect them to hold a stronger handle to weigh in on the potential threats of a model, thus giving themselves the ability to slow down future rollouts of new models. In fact, OpenAI's Sam Altman, an ardent liberalist when it comes to AI innovation, went on record saying he wished the race slowed down. According to the NYT, representatives of Anthropic, OpenAI, Microsoft, Meta, Google and Nvidia were among those who met White House representatives last Tuesday to gauge their views around regulating AI. Of these companies, Nvidia, Meta and Google make open-source models as do DeepSeek, Moonshot AI and the Mira Murati-led Thinking Machine Labs. When the White House announced their intention to create a regulatory framework, it was assumed that the Chinese models could be at the forefront, given the fact that their recent models performed almost as well as their American counterparts. Maybe, Washington realised that getting to regulate Chinese models via this review process may not work. Several experts hold the view that a regulatory framework around AI has to evolve into a global model and not just limit itself to a country and its perceived rivals. The focus should be on creating guardrails around advanced AI that evolves with the technology itself and brings governments of other countries to a negotiating table. White House Spokeswoman Liz Huston said in a statement that the voluntary framework aims to advance the Trump Administration's America Frist cybersecurity strategy by strengthening "our national security" and cementing American AI dominance. And companies that choose to collaborate through this framework are putting American innovation and security first. Given that President Trump hasn't really cared to form global opinion on anything in his second term, having instead used tariff barriers time and again, we wonder how such a framework would pan out into a global regulation. Most likely not. Which means enterprises could still favour cheaper Chinese models in the United States. We can only wait and watch!