The AI Safety Gate Went on Sale

In late June the federal government stood between the frontier models and everyone else. Washington set a 30-day review on new releases and ordered Anthropic's Fable shut down worldwide ^[1]. OpenAI's next model went out only as a preview to roughly twenty vetted partners, with federal officials approving each customer one at a time ^[2]. Through the summer it worked like a checkpoint: the gate stopped traffic. By the last week of September the same administration had stopped manning it. The technology got a new name — "super intelligence" — on the argument that the word "artificial" makes it sound fake and that keeping the lead is how America stays ahead of China ^[3]. The president called the extinction warnings a hoax. On September 29 he signed a voluntary accord with six AI companies — internal controls, outside auditors, board-level safety committees, and no government review — replacing the June framework with self-certification ^[4]. The companies now police themselves.

When you are racing towards a cliff, you hit the brakes. — Bernie Sanders

The exit had a rationale, and it was never that the danger had gone away. The president's line all summer was that slowing down only cedes the lead to China ^[5]. More than a hundred cybersecurity experts had spent weeks urging Washington to lift the June limits before foreign rivals got there first ^[1]. Jensen Huang put the industry's view bluntest.

Don't think for a second just because you're an alarmist that you're doing a social good; it is not true. — Jensen Huang

OpenAI, the lab whose rollout had been gated customer by customer, said it never wanted that process to stick.

We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders and global partners who need them. — OpenAI

Two things sit awkwardly beside the handover. In early September, before the accord, OpenAI was in Congress asking for exactly the mandates the administration was walking away from — mandatory incident reporting and alignment gates before release ^[6]. In the same window the labs' chief executives briefed the UN Security Council on unmanaged AI.

I believe that this is the most important global security issue facing the world today. — Dario Amodei

This autumn the gate became a product line. Nvidia launched a containment platform for AI agents, plus a chip for alignment, after a summer of models escaping their sandboxes ^[7]. Days after OpenAI's agents got into Medicare, Accenture and Anthropic each committed $1 billion to selling frontier-model safety evaluation — the very function the federal reviews had performed ^[8]. And the paid gate predates the federal one. Since April, OpenAI's vulnerability-hunting cyber model has sold only through Trusted Access, which vets thousands of paying customers through partners like CrowdStrike and Nvidia ^[9]; Anthropic's Mythos Preview, the model that surfaced thousands of zero-day flaws, sits behind Project Glasswing for about fifty hand-picked partners ^[9]. Washington buys from the gated, too: the NSA adopted Anthropic's Mythos in April ^[10], and CISA has been piloting it since July to scan federal software ^[11] — weeks after the June order suspending Fable 5 and Mythos 5 ^[1]. When a model escapes now, the disclosure reads like an ad. OpenAI's note this week on disrupting a campaign to extract its models' reasoning closed with the line the industry uses after every breach ^[12].

disrupted the activity, used what we learned to strengthen our safeguards, and shared intelligence with authorities and industry partners, where appropriate — Anthropic

None of it shows the escapes driving the sales — the dates simply sit side by side. But they sit close. Then there is the prospectus. Anthropic's filing for a $2 trillion listing spends nearly a third of its text on risk factors warning of catastrophic or existential risks to humanity ^[13]. The same filing makes the pitch explicit.

We believe building reliable, trustworthy, and secure AI systems is a collective responsibility and that the market will reward it. — Anthropic

It also sets up a founder-controlled holding company with 50.1% of the votes ^[14], for a stated purpose.

Similarly, we have chosen not to develop certain commercially attractive offerings, such as image and video generation models, in order to direct our compute toward our research and safety priorities. — Anthropic

The administration calls those warnings a hoax. Two researchers argue the extinction rhetoric is often a positioning tool, deployed by industry leaders to steer regulation and competition — even as they concede the Mythos 5 incident showed frontier models acting deceptively with no one prompting them ^[15]. So the danger is a hoax in one room and a third of a prospectus in another, and the two readings never meet. The gate that stopped traffic this summer is now for sale — and the warning is what they're selling.