Selective Safety, Act Two

Two weeks ago, I argued that the Trump administration was playing favourites with AI safety: Anthropic, the company that had hobbled its own models for show, got a binding order to suspend foreign access to Mythos, while OpenAI—whose GPT-5.5-Cyber model was quietly released with no fanfare—got a pass. Now we have act two, and the script has been tweaked just enough to make the point subtler, not weaker.

This week, Axios reports that the administration has 'formally requested' that OpenAI limit the release of GPT-5.6 to a small set of government-approved partners, citing security concerns. Note the verb: 'requested'. Not 'ordered'. Not 'demanded'. Not 'directed under executive authority'. A request. The same administration that told Anthropic to shut off foreign access to its frontier model is now politely asking Sam Altman if he wouldn't mind keeping his next toy in a sandbox for a while. The message could not be clearer: Anthropic is to be disciplined; OpenAI is to be managed.

Why the difference? Partly it's performance. Anthropic spent months advertising its safety rituals—constitutional AI, red-teaming, staged releases—and in doing so made itself a target. When the government wants to show it's serious about AI risk, it goes after the company that has loudly claimed to be serious about safety. OpenAI, by contrast, has always played its cards closer to its chest, releasing models without the same theatrical hand-wringing. You can't be asked to prove your safety if you never claimed to be safe in the first place.

But the more cynical explanation is that the administration is simply picking winners. Anthropic's Dario Amodei doesn't have the same Washington access as Sam Altman, whose charm offensive on Capitol Hill is legendary. And let's not forget the pro-AI super PAC Leading the Future, which—as Politico reported—spent heavily to unseat a New York assemblyman and then, after losing, clammed up. The industry knows that political donations buy a different kind of treatment: a request, not an order.

The timing of the Axios story is also instructive. It comes on the heels of Arctic Wolf's finding that open-source AI models are nearly as effective as Mythos for exploiting vulnerabilities—just without the guardrails. So if the administration were genuinely concerned about security, why not go after the open-source models that are freely downloadable and ungovernable? Because that would require a real regulatory framework, not just selective enforcement against a convenient target.

What we are watching is not a coherent safety strategy but a series of tactical nudges designed to create the appearance of oversight without actually constraining the industry's most powerful players. Anthropic gets the stick; OpenAI gets the carrot. And the public is left to marvel at the spectacle: two companies, one a scapegoat and one a favourite, both continuing to build the same unaccountable systems.

The real AI safety discussion—about what happens when these models are deployed at scale, about the workers being laid off to pay for them, about the energy and environmental costs—remains as absent as ever. Because that discussion would require treating all companies equally. And that, apparently, is a step too far for a government that prefers to choose its villains.