Yoshua Bengio is one of the three people who can legitimately claim to have invented modern AI. When he speaks, the room usually quiets down. This week, he used that authority to make a point that does not fit neatly into either camp of the AI debate.

Bengio's complaint is not that Anthropic built Mythos, a model capable enough at offensive cybersecurity that the UK AI Security Institute called it the most capable cyber-offensive system it has ever evaluated. His complaint is that Anthropic alone gets to decide who can use it.

A handful of US companies. A handful of US agencies. Allies not on the list yet. The rest of the world waiting for an invitation. This is not a hypothetical about superintelligence ten years from now. This is happening this quarter, inside a program called Project Glasswing.

The AI Safety Argument Cuts Both Ways

Anthropic's reasoning is internally coherent. Mythos is dangerous enough that releasing it openly would hand a weapon to every state-aligned hacking group on the planet. Selective release to trusted partners buys time for defenders to patch the vulnerabilities Mythos uncovers before attackers find and weaponize the same ones. That is a reasonable position, and it is the one Anthropic has taken.

Bengio is pointing at the flip side. If you accept that a model is too dangerous for open release, you have just conceded that a single private company is now making foreign policy by other means. "Trusted partners" is a term that has traditionally been the domain of states, negotiated through treaties and alliances. Anthropic's trust list is negotiated through sales calls.

Who Actually Polices This?

Bengio's proposed answer is an agency. An actual government body with technical staff whose job is to decide which frontier models get deployed where. That is not a fashionable answer in 2026. The Trump administration is trying to preempt state AI laws and wants the lightest possible federal touch. Congress is not close to passing anything substantive on frontier model governance.

In that vacuum, Anthropic is doing the thing a government would normally do. So is OpenAI. So is Google DeepMind. The labs are not acting maliciously. They are responding rationally to a policy environment that refuses to provide rules, and then the labs get blamed for making up their own.

Where This Ends

There are two endings to this story. In one, governments get serious about frontier model oversight, and access decisions move out of private boardrooms and into a multilateral process that can be audited. In the other, the current pattern hardens into precedent, and the geopolitical balance of AI shifts based on which labs decide which partners are trustworthy on any given Tuesday.

Bengio is warning that we are drifting toward the second ending by default. Nothing about the Trump White House's approach, or Congress's inaction, or the labs' commercial incentives, is going to reverse that drift. It will take a deliberate choice to build something new. And nobody in power currently wants to make that choice.

So we will get the second ending. And when something bad eventually happens, and it will, the retrospective will blame Anthropic for being the gatekeeper. That will be wrong. Anthropic filled a gap the state refused to fill. The correct target of blame is the state that refused.

First reported by The Times of India and NewsBytes, based on Bengio's public commentary this week.