It’s Out There—But Who Said “Go”?
OpenAI just unleashed Sol into the wild. If you’ve been watching frontier AI models, that sounds odd—and it should be. Sol sits on the same performance tier as Anthropic’s Fable, a model so volatile the White House temporarily barred its broader use. And yet—no public safety review, no published evaluation report, no independent audit summary.
You’d think that for something as explosive as a frontier model capable of novel autonomous reasoning, the government would lay down some ground rules before letting it loose on millions of users. But as it turns out, nobody’s really sure what the rules are—or who gets to set them. OpenAI says it consulted with top federal officials, but the details remain locked up. When TechCrunch asked how exactly Sol was cleared, their answer boiled down to “look at the safety card and external evaluations” (we’ll get there), not a transparent, public process.
This isn’t just about Sol. It’s about the growing disconnect between where AI capabilities stand today and how—or whether—we, as a society, sign off on their release. It’s the same frustration Anthropic felt last year when Fable’s rollout got derailed behind closed doors. And it’s a growing unease among safety researchers, former government advisors, and even industry insiders who worry that the only criteria for approval is whether the right person at the Treasury Department got a coffee with Sam Altman.
The “Just Ask” Process: Or, How No One Asked
Mina Narayanan, a senior research analyst at Georgetown’s Center for Security and Emerging Technology, laid it out bluntly in a TechCrunch interview: “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.”
That’s the core of the problem, right there—not malice, but opacity. Anthropic claimed it had consulted with the government before rolling out Fable, developed jailbreak classifiers, and implemented defense-in-depth strategies. But when the dust settled, Fable got restricted—not because it failed a test, but because nobody knew what that test was. OpenAI’s Sol feels like a rerun, with the same missing cast: transparency, public accountability, and even internal clarity at OpenAI itself.
Dean W. Ball, a former Trump policy advisor turned OpenAI insider, wrote in a recent newsletter: “nobody knows what the requirements are to get licensed.” And he’s not just being diplomatic. Andy Konwinski, the computer scientist behind Databricks and co-founder of Perplexity and Laude Institute, put it differently: “It’s existentially a problem…who gatekeeps and decides on permissions?”
There’s something deeply unsettling about fronting a new superpower without a driver’s license for it. The U.S. government wants AI innovation to happen, sure—but it also wants it happening somewhere visible, with guardrails. Right now? Those guardrails aren’t bolted down anywhere.
The White House last month published an executive order laying out a roadmap, but it explicitly ruled out what most people assumed would be inevitable: an FDA for AI. That isn’t just a semantic nitpick—it’s a structural gap. Without a dedicated agency, oversight gets parceled out across departments, each with its own mandate, budget constraints, and political agenda. Sriram Krishnan, who served as a senior advisor for AI in the White House until this past June, told the Financial Times bluntly that no consensus has emerged on which models require scrutiny—or who gets to decide that.
The Paper Trail That Isn’t There
OpenAI CEO Sam Altman told CNBC that officials like Secretary of Commerce Howard Lutnick, Treasury Secretary Scott Bessent, and Cyber Director Sean Cairncross were consulted before Sol’s release. But TechCrunch couldn’t get any details, and OpenAI declined to share them publicly.
What they did point to? The model’s safety card—along with external evaluations by U.K. AISI, SecureBio, and Irregular. That’s not nothing; it shows OpenAI knows third-party review matters. But it also sidesteps the harder question: what did those evaluations do? Were there adversarial prompts tested? Did they stress-test jailbreak resistance? Did any of those organizations have the access and time to run meaningful red-teaming exercises?
The answer, from outside the room? We don’t know. And yet Sol launched anyway.
What’s especially revealing is OpenAI’s own admission: in a late June blog post, they wrote they “don’t believe this kind of government access process should become the long-term default.” Translation: we did what we had to do this time, but only because someone told us to. Long term? We’ll run the lab, and your oversight can catch up.
That kind of confidence is only possible if you already control the infrastructure, talent pool, and narrative. And—here’s where it gets icky—when one of your top executives is the largest known donor to a sitting president’s mid-term political operation.
Greg Brockman, OpenAI’s president, reportedly bankrolled more than anyone else for Trump’s mid-term efforts. Sam Altman? Ball wrote that he’d offered up to 5% of OpenAI’s equity for the so-called “Trump Accounts.” Coincidence? Maybe. But when your safety process depends on who you grab coffee with at the Treasury, it’s hard to argue that merit or risk assessment is driving the decision.
The Ghost of Fable—and Why It Haunts Now
Anthropic’s Fable was pulled from wider access not because it failed basic safety checks, but because the government withheld permission—partly over actual jailbreak concerns and partly over personality clashes between Anthropic’s leadership and the current administration.
OpenAI, by contrast, appears to have sailed through without public hurdles. That tells me two things: one, OpenAI has built a much closer relationship with this administration; two, the government’s internal process is so muddled that “approval” now means “someone high enough in the chain nodded.”
Fable’s brief ban taught Anthropic to prepare for worst-case scenarios. OpenAI didn’t need to—it had a different shortcut: inside access.
But inside access isn’t a safety framework. It’s a luxury. And if you don’t have it, your model gets flagged as high-risk for arbitrary reasons (like personality friction). If you do have it? Your model ships without scrutiny. That’s not resilience—that’s rent-seeking with neural nets.
The real casualty here is public trust. Remzi Arpaci-Dusseau, a computer science professor at University of Wisconsin-Madison, said it best at the Open Frontier conference last week: “There’s not a sense that responsible people are driving forward these changes.”
He didn’t mean “responsible” as in “well-meaning.” He meant qualified—safety researchers, alignment experts, interpretability specialists, data teams—not just executives and political appointees.
The Commons Option: What Actually Works
Andy Konwinski has been pushing a different model for years—one based on the NIH, FDA national labs: open commons, where government, academia, and private players co-develop safety standards before launch.
He’s right. Locking safety review inside a few tech giants—even with “external” evaluations—means you’re relying on the same folks who built the model to prove it doesn’t detonate. That’s like asking a mechanic to certifying their own engine.
His proposal? Licensed third-party auditing organizations, sanctioned by the government, that would evaluate frontier labs’ safety protocols. Think: independent red-teamers with academic ties and real access—not consultancies paid by the lab itself.
Or, better yet, a new institutional form—a “National AI Lab” or focused research organization—where disinterested experts from universities and nonprofits can evaluate frontier models before they go public.
The incentives matter. Right now, Konwinski points out, even “good-intentioned” companies face fiduciary pressures to recoup training costs quickly and stay ahead of the competition. Legal obligations override risk-aversion because the board (and shareholders) demand it.
That’s not a criticism of OpenAI alone. It’s a critique of the whole incentive structure we’ve allowed to take root. AI isn’t being developed in a vacuum; it’s being race-casted. And racing without guardrails looks heroic on the highlight reel, but terrible when you crash.
David Siegel put it starkly at that same conference: imagine a world where “a small number of firms control the technology; the government, in their secretive laboratories, is evaluating whether or not the technology is suitable for use; and the general public and scientific community doesn’t really have any access to any of that stuff.”
He asked us to imagine it. But we don’t have to—Sol is already live.
The Deadline That Isn’t Coming
The executive order mandates six cabinet agencies to finalize an evaluation process by early August. But “finalizing” doesn’t mean “implementing.” And even if they meet that date, who knows what the rules will look like—especially when the agencies involved include Defense, Homeland Security, Commerce, Education, Health and Human Services, and Energy.
Each has its own view on AI: Defense wants offensive capabilities; Homeland Security worries about border and infrastructure security; Commerce focuses on standards and trade. Getting them to agree on a baseline—let alone enforce it—is heroic.
Meanwhile, the labs keep pushing. And if history’s any guide, they’ll push right up to—and past—the deadline, betting that public pressure won’t catch up before launch day.
This isn’t sustainable. A model’s readiness shouldn’t hinge on whether the right official is in the loop, or whether someone at Treasury has a soft spot for Altman’s pitch. It should hinge on repeatable,公开 audits, published methodologies, and a clear threshold for what counts as “safe enough.”
We don’t have that. And Sol’s release is a wake-up call—not because it’s dangerous, but because its approval process was almost entirely opaque.
There’s time to fix this. But the clock starts now—before Sol 2 drops with even less oversight than its predecessor.