What Led the Government to Consider OpenAI’s Frontier Model Safe for Release?

OpenAI is launching its newest advanced LLM, Sol, for public accessibility. Sol is believed to be at least on par with Anthropic’s Fable, a model whose capability raised enough concern in the White House that it was briefly restricted from public use.

So, what facilitated the approval of these models for release? The concise answer: The details remain ambiguous.

“To be honest, I lack insight into those specific processes, so I cannot assess their adequacy with confidence,” said Mina Narayanan, a senior research analyst at Georgetown’s Center for Security and Emerging Technologies, to TechCrunch. “Anthropic mentioned engaging with the government and developing a classifier to detect jailbreak attempts, along with employing strategies to avert future ones, yet the particulars of the interaction between the government and entities like Anthropic and OpenAI are not clearly defined.”

Dean W. Ball, a former policy advisor during the Trump administration now affiliated with OpenAI, indicated in his newsletter last month that “nobody comprehends what the licensing requirements are.”

Andy Konwinski, a computer scientist and co-founder of Databricks, Perplexity, and the Laude Institute, expressed he has yet to meet anyone who understands the process, even within frontier labs. “This is a critical issue,” he remarked to TechCrunch. “Whether for safety or otherwise, it ultimately comes down to who has the authority to make decisions—who is the gatekeeper and controls the permissions?”

Eighteen months into the Trump administration, clarity on the way forward remains elusive, partly due to industry figures’ involvement in policymaking. Last month, after weeks of internal conflict, an executive order was released outlining a framework for evaluating frontier models, but specific details are still forthcoming. “There will not be an FDA for AI,” stated Sriram Krishnan, a former partner at Andreessen Horowitz who served as a senior AI advisor in the White House until last month, in a statement to the Financial Times.

There is currently no agreement on which model types require government oversight or which agencies should perform those evaluations. At present, the Department of Commerce’s Center for AI Standards and Innovation seems to be spearheading the effort, but the executive order mandates six cabinet agencies to finalize a process by early August. Meanwhile, what has emerged is, at best, a fragmented approach.

OpenAI CEO Sam Altman disclosed on CNBC that the process included discussions with figures such as Secretary of Commerce Howard Lutnick, Secretary of the Treasury Scott Bessent, and US national cyber director Sean Cairncross; however, the identities of the experts who assessed the models remain undisclosed, along with their evaluation methodology. OpenAI opted not to reveal details about the government’s evaluation process to TechCrunch but referenced findings from various external reviews conducted by organizations like UK AISI, SecureBio, and Irregular in the recent model’s safety report.

Similar to Anthropic’s Fable rollout, OpenAI provided a preview of the model to government officials and select users prior to the wider release; however, the identities of those users and the criteria for their selection remain undisclosed. In a blog post from late June, the company stated, “we don’t believe this type of government access process should become the standard long-term practice,” affirming a commitment to collaborate with the government on an alternative method moving forward.

However, the context of these discussions includes Altman reportedly offering up to 5% equity in OpenAI for the administration’s “Trump Accounts,” coupled with OpenAI president Greg Brockman being the largest publicly known donor to Trump’s mid-term political campaign. This raises concerns about the impact of these activities on the government’s apparently lenient stance towards regulating Sol.

In contrast, Anthropic’s Fable was briefly restricted from broader access when the US government barred its use by foreign nationals, partially due to valid concerns about users jailbreaking the model for hacking capabilities and partially due to tensions between Anthropic and the Trump administration. The possibility of an export ban may have also motivated OpenAI to comply more fully with the government’s unspecified requests.

From an industry perspective, a laissez-faire regulatory approach may seem attractive, but one that hinges on personal relationships with administration officials brings uncertainty and potential conflicts of interest.

Konwinski communicated to TechCrunch his worry that genuine experts in this field—“safety researchers, alignment researchers, interpretability researchers, alongside those dealing with data and various other stacks”—are not adequately involved in the model release process.

Konwinski proposes that an “open commons” model would effectively balance safety and innovation. He cites existing models like the FDA, NIH, or national labs, which unite researchers, government officials, and private entities to achieve consensus on safety matters.

Part of this dilemma originates from the capitalist incentives that have historically driven AI researchers—an issue that has played out in court during Elon Musk’s lawsuit challenging OpenAI’s corporate structure. Ball highlights that the AI industry’s nature necessitates companies to recuperate a substantial portion of their training costs soon after launching their models to remain competitive.

“Even with noble intentions, there are clear legal obligations and fiduciary responsibilities embedded in the operating procedures,” Konwinski stresses.

In his post, Ball argued that progressing forward will necessitate third-party auditing organizations, sanctioned by the government, to evaluate frontier labs’ safety protocols. Konwinski is also optimistic about the potential of new institutional structures, such as focused research organizations, to enhance access to and assessment of frontier models by impartial experts from academia and the non-profit sector.

For now, the secrecy surrounding AI development shows no signs of abating, which may lead to political challenges for an industry increasingly viewed with suspicion by the public. “There’s a lack of confidence that responsible individuals are directing these advancements,” commented University of Wisconsin-Madison computer science professor Remzi Arpaci-Dusseau last week at the Open Frontier conference.

During the same event, David Siegel, the computer scientist who founded Two Sigma, one of the most successful quantitative hedge funds, urged attendees to “envision a concerning scenario, [where] a limited number of firms control the technology; the government, in secretive labs, evaluates the technology’s suitability for use; and the general public and scientific community lack access to any of that information.”

It seems we don’t need to merely envision it.