How Did the Government Decide OpenAI’s Frontier Model Was Safe to Release?

OpenAI has begun rolling out Sol, its latest and most powerful large language model, making it widely available to the public. The system is considered at least as capable as Anthropic’s Fable, a model whose advanced features and potential risks raised enough concern within the White House that it was temporarily restricted from general access. Yet despite the high stakes, the question of exactly how regulators determined Sol was safe enough for release remains shrouded in uncertainty. As experts, industry leaders, and observers point out, the process appears to be ad hoc, opaque, and lacking in clear rules or public accountability.

A Process Lacking Clarity

For anyone seeking to understand how such high‑stakes decisions are made, the answer is remarkably vague. “Frankly, I don’t have visibility into those exact processes, so yes, I don’t feel like I have enough information to say whether they’re adequate or not,” said Mina Narayanan, senior research analyst at Georgetown University’s Center for Security and Emerging Technology. While Anthropic has stated it held discussions with officials and developed tools to block attempts to bypass safety controls, the details of those conversations, the criteria applied, and the evidence reviewed have never been fully laid out.

The confusion extends even to those inside the industry. Dean W. Ball, a former Trump administration policy advisor now working at OpenAI, wrote in his newsletter last month: “Nobody knows what the requirements are to get licensed.” Andy Konwinski, a computer scientist and co‑founder of Databricks, Perplexity, and the Laude Institute, put it more bluntly: “I’ve never spoken to anyone who understands the process, even employees at the frontier labs themselves. It’s existentially a problem. Safety or not, it’s about who has the power to make decisions wh, who gatekeeps, and decides on permissions?”

Regulatory Framework Still in Flux

Eighteen months into the current administration, the United States still lacks a formal, agreed‑upon system for evaluating and approving advanced AI models. Following weeks of internal debate, an executive order was finally released last month outlining a general roadmap, but it left almost all the key details undefined. One of the few clear statements came from Sriram Krishnan, a former senior White House AI advisor, who told the Financial Times: “There will not be an FDA for AI.”

There is still no consensus on which models require government oversight, what level of capability triggers review, or which federal agency should lead the assessments. For now, the Department of Commerce’s Center for AI Standards and Innovation has taken the lead, but the executive order tasks six different cabinet departments with finalizing a unified process by early August. Until then, decisions are being made on a case‑by‑case basis, with little consistency or transparency.

OpenAI CEO Sam Altman has said the company held discussions with senior officials, including Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent, and National Cyber Director Sean Cairncross. But the company has declined to share who specifically tested the model, what safety benchmarks were used, or how the final determination was reached. Instead, it has pointed to external assessments from organizations such as the UK’s AI Safety Institute, SecureBio, and Irregular, which are listed in the model’s published safety documentation.

As Anthropic did before releasing Fable, OpenAI previewed Sol to government officials and a small group of select users ahead of its public launch, but the identities of those users and the scope of their feedback remain undisclosed. In a late June blog post, the company noted it “does not believe this kind of government access process should become the long‑term default,” adding that it hopes to work with authorities to develop a different approach moving forward.

Questions of Influence and Fairness

Adding to the uncertainty are questions about the relationship between OpenAI and the administration. Reports have surfaced that Altman has offered as much as 5% of OpenAI’s equity to support the White House’s so‑called “Trump Accounts” initiative, while OpenAI President Greg Brockman is listed as the largest publicly known donor to the president’s political operations. These connections have led critics to wonder whether the relatively light‑touch treatment Sol has received is based on safety evidence or on personal and financial ties.

The contrast with Anthropic’s experience is striking. Fable was temporarily restricted, and foreign nationals were barred from using it, partly due to concerns that bad actors could bypass safeguards to access hacking and cybersecurity tools, and partly due to reported tensions between the company’s leadership and the administration. The threat of similar restrictions, industry sources suggest, may have encouraged OpenAI to be more cooperative with officials even if the terms of that cooperation remain secret.

Calls for a More Transparent System

From a broader perspective, the current approach creates more problems than it solves. While a hands‑off regulatory style may benefit companies in the short term, relying on informal negotiations and personal connections creates uncertainty, uneven treatment, and the risk of regulatory capture.

Konwinski argues that true expertise is being sidelined in the process. “Safety researchers, alignment researchers, interpretability specialists, data scientists, and engineers across the entire technology stack, these are the people who should be guiding these decisions, not just executives or political appointees,” he said. Instead, he advocates for a model similar to the U.S. Food and Drug Administration or the National Institutes of Health, independent bodies that bring together scientists, regulators, and industry to set clear standards, conduct independent reviews, and publish their findings publicly.

There are also economic pressures at play. As Ball has noted, the business model for advanced AI requires companies to recoup massive training costs quickly, while staying ahead of competitors, creating an inherent incentive to release models as soon as possible, even if safeguards are not fully proven. “Even if their intentions are good, there are clear legal and fiduciary obligations built into their operations that push them toward speed over caution,” Konwinski added.

Proposals to fix this range from establishing government‑licensed third‑party auditors to creating independent research organizations with access to frontier models, so that evaluations can be conducted by experts without commercial or political conflicts of interest.

Growing Public Skepticism

As the secrecy around model approvals continues, it risks deepening public mistrust in the technology. “There’s not a widespread sense that responsible people are driving forward these changes,” said Remzi Arpaci‑Dusseau, a computer science professor at the University of Wisconsin‑Madison, speaking at the recent Open Frontier conference.

At the same event, David Siegel, founder of the quantitative investment firm Two Sigma, warned of the risks of concentrated control: “Imagine a situation which I think would be very bad where a small number of firms control the technology; the government, in its secretive laboratories, decides whether it’s safe; and the broader scientific community and public have no access to the details.”

As the launch of Sol shows, however, that scenario is no longer just hypothetical. For now, the decisions that shape the future of powerful AI systems are being made behind closed doors, with few rules, limited oversight, and little way for anyone outside the room to know whether the choices being made are truly in the public interest.