Sol's Secret Sauce: Nobody Knows What Safety Looks Like

Last week I wrote about the GPT-5.6 pardon, the Trump administration's quiet lifting of restrictions on OpenAI behind closed doors. I called it selective safety. This week, the sequel arrives on schedule, and it's worse. OpenAI has released Sol, its new flagship frontier model, after receiving government approval. But ask how that approval was obtained, and the answer appears to be: nobody knows.

TechCrunch reports that former Trump policy adviser Ball stated flatly that 'nobody knows what the requirements are to get licensed' and that it is unclear who tested the models or how the evaluation was conducted. This is not a minor oversight. This is the regulatory equivalent of a police officer handing you a driving licence after you wave at them from behind the wheel. The government has created a licensing system for frontier AI models — ostensibly to prevent catastrophic risks — and then refused to disclose the criteria, the testers, or the methodology. It is a safety theatre of the highest order.

Let me be precise about what we don't know. We don't know which organisation ran the red-teaming. We don't know what benchmarks were used. We don't know whether the model was tested for bioweapons design, cyberattack capabilities, or autonomous replication. We don't know if the evaluation lasted three weeks or three months. We don't know if the evaluators found anything worrying and, if so, what mitigations were demanded. The only thing we know with certainty is that the model was approved, because OpenAI is now selling it.

The convenient opacity extends to the company itself. CEO Sam Altman told CNBC that Sol is 54% more token efficient on agentic coding tasks — a precise, measurable claim. But on safety, the company offers only vague assurances. The contrast is instructive: when there is a commercial advantage to be gained, numbers appear. When there is a regulatory obligation to be met, silence follows. And the government, which is supposed to hold the line, abets the silence.

This is not a partisan point. The previous administration began the licensing framework; this one is implementing it. Neither side has shown any appetite for genuine transparency, because genuine transparency would require either blocking a major product release — politically unpopular — or admitting that the whole apparatus is a paper tiger. It is far easier to nod, stamp the form, and move on. The result is a system that provides cover for companies without providing safety for the public.

You might argue that some details must remain confidential to avoid helping bad actors. I will grant you the theoretical point. But there is a difference between protecting operational security and hiding everything. The government could release the names of the testing organisations without revealing their internal methods. It could publish a summary of the threat scenarios evaluated without specifying every parameter. It could confirm that specific red-teaming results were addressed without detailing the fixes. None of this has happened. The black box is complete.

What makes this especially galling is that the entire premise of the licensing regime was to prevent exactly the kind of unaccountable release we just witnessed. When the AI Safety Institute was established, its proponents promised rigorous, transparent evaluation. Instead, we have a rubber stamp in a different font. The lesson for the industry is unmistakable: the safety process is a hoop to jump through, not a barrier to overcome. And the lesson for the public is equally clear: you are trusting your safety to a process that its own architect admits is unknowable.

OpenAI will sell millions of Sol subscriptions. Companies will integrate it into critical infrastructure. And nobody — not the government, not the company, not any independent auditor — can explain with confidence what risks were considered before that happened. The 5.6 pardon was a giveaway. This is a blank cheque.