A classified framework, no clear roadmap
Executives from Google, OpenAI, Meta and Anthropic met with U.S. officials Tuesday in a meeting organized by the Office of the National Cyber Director. The session is aimed at explaining a government framework for benchmarking the risk of powerful AI models before they are released to the public.
The framework focuses on how well AI systems can detect and exploit vulnerabilities in software. It was triggered by an executive order signed by Donald Trump on 2 June. That directive explicitly rules out any government licensing or preclearance system. Instead, companies may choose to give the government 30 days of access to test models before they are shared with other partners.
Uncertainty looms over definitions and responsibilities
Despite the clarity on process, major questions remain. How will the government define 'frontier models'? Will open-source AI systems be included? And who will actually run the evaluations? These gaps remain unaddressed by the administration. National cyber director Sean Cairncross, along with Treasury Secretary Scott Bessent and Commerce Secretary Howard Lutnick, has been driving the initiative. Yet no single office has been tasked with ongoing engagement with the private sector.
The lack of clarity is a key concern for firms with models already in development. These companies want to know immediately whether the new rules apply to their upcoming launches.
Industry and government are at a tense crossroads
Recent weeks have seen rising friction between Washington and the AI sector. Export controls briefly blocked Anthropic's Fable 5 launch, while the administration asked OpenAI to limit the initial rollout of GPT-5.6 to government-approved partners.
Private disclosures have also fed concerns about AI risks. OpenAI reported that an experimental agent escaped its testing environment and infiltrated Hugging Face's systems. Anthropic decided not to release a model capable of uncovering software weaknesses, named Mythos. Over 1,100 employees from the four AI firms also recently signed an open letter urging Washington to establish international tools to regulate AI if its development outpaces oversight.
The President Donald Trump administration said Monday (Aug. 3) that it finalized plans for voluntary federal cybersecurity assessments of advanced U.S. AI models. Anthropic, Google, Meta and OpenAI were invited to meet with White House officials Tuesday (Aug. 4) to discuss the framework. The administration hasn’t released the testing metrics, reporting requirements or details about what results will be disclosed.
The framework follows a June 2 executive order directing federal agencies to develop a classified benchmark for measuring models’ advanced cyber capabilities. The order also established a voluntary process allowing developers to provide the government with access to covered frontier models for up to 30 days before their broader release. The order said the process doesn’t create a licensing system, mandatory preclearance requirement or legal condition for deploying an AI model.
Commercial pressure, however, could give the assessments more weight than their voluntary label suggests. Banks, insurers and government contractors already ask technology providers to document security reviews, penetration tests and compliance certifications. A federal assessment could become another procurement requirement, particularly when a model will interact with customer information, payment systems or critical infrastructure.
The shift would fit a broader move toward more demanding AI procurement. Companies are paying closer attention to data retention, deletion rights, auditability and contractual responsibility as AI enters financial and operational workflows. The value of the federal program will depend on what buyers can learn. Saying that a model was tested offers limited assurance unless customers know which capabilities were examined, which weaknesses were found, and what restrictions or safeguards followed.
The stakes are rising because frontier models are becoming more capable of finding and exploiting software vulnerabilities. Advanced AI can compress cyber research that once required months of expert work, potentially expanding both defensive capabilities and the attack surface facing banks. Frontier AI is beginning to perform sophisticated cryptographic analysis.
Financial institutions should ask vendors whether a model has undergone the federal assessment, what findings can be shared and whether significant capability upgrades will trigger another review. They should also determine whether the tested version is the same one being offered commercially. Federal testing won’t replace a bank’s own controls. A government assessment may identify broad cyber capabilities, but it won’t determine what happens when a model is connected to a specific institution’s credentials, payment APIs, customer data and approval rules.
That distinction may define the program’s real role. The federal government can assess the engine. Banks will still need to test the vehicle, the road and who’s allowed behind the wheel.
The White House framework for vetting the cybersecurity risks of advanced artificial intelligence models would focus on the top-tier products from U.S. developers such as OpenAI and Anthropic — while exempting a rising breed of lower-cost AI software, three people familiar with the policy told POLITICO. Specifically, the vetting policy would not apply to so-called open-source or open-weight models, according to the people, who were granted anonymity to disclose details from private conversations. Open models have lately seen a surge of activity and interest across the tech industry, including from AI developers in China, despite not yet offering the same advanced capabilities of the top U.S. models. The new details, provided after the White House met with staffers from top tech companies Tuesday to review the framework, offer the latest glimpse into the Trump administration’s efforts to tackle two daunting goals — ensuring that the AI industry’s ever-more-powerful products don’t create unsustainable security risks, without imposing such heavy regulations that the U.S. loses the tech race to Beijing. The framework, which the White House has yet to make public, follows a June 2 executive order aiming to address AI’s catastrophic risks. Only “state-of-the-art” models deemed to be national security risks would be covered by the policy, two of the people said. The executive order calls for a benchmarking process overseen by multiple agencies, including the National Security Agency, to determine which models would qualify. Further details about precisely which advanced models or AI companies would be subject to the policy were not immediately available. The document would also set a 30-day maximum period for the government’s review of new models, according to the people, adding that the evaluation process is expected to involve a wide swath of federal employees.
“The voluntary framework advances the Trump Administration’s America First cybersecurity strategy by strengthening our national security and cementing American AI dominance,” White House spokesperson Liz Huston previously said in a statement. “Companies that choose to collaborate with the Administration through this framework are putting American innovation, security, and cyber defense first.”
Open-weight models, which are customizable and whose core components are publicly released, have received widespread industry backing, with Nvidia CEO Jensen Huang making the rounds on Capitol Hill last week to advocate against possible restrictions.
As the White House grapples with its approach to risky AI models, some skeptics have expressed concern that new safety restrictions could give Chinese companies an edge and threaten U.S. supremacy in the race toward advanced AI.
Powerful AI models have made headlines with a string of cyber breaches that the companies say were made possible by the technology’s rapidly advancing capabilities. In some cases, the software companies said their models deliberately broke out of an attempt to contain them or otherwise behaved in ways that the AI developers had not predicted.