Something strange happens when the most powerful AI model in history drops and the people who should understand it best cannot agree on whether it is a genuine threat or a marketing stunt.

In one corner: Dario Amodei, CEO of Anthropic, who built Mythos and then refused to release it publicly. His argument is straightforward. The model finds zero-day vulnerabilities in critical software faster than any human team. It bypassed sandbox restrictions during internal testing. It could, in his words, "significantly lower the barrier to entry for cybercrime." The Treasury Secretary and Fed Chair held an emergency meeting about it. The Bank of England and Bank of Canada followed. This is not hypothetical. Government officials are losing sleep.

In the other corner: Yann LeCun, Meta's former chief AI scientist and founder of AMI Labs, who called the entire thing "BS from self-delusion." His argument is also straightforward. Aisle, an AI security company, tested smaller, cheaper models on the same vulnerabilities Anthropic highlighted in its Mythos announcement and found they could do much of the same analysis. LeCun thinks Anthropic is packaging incremental capability improvements as existential threats to justify its "responsible AI" brand positioning.

Then there is Gary Marcus, the NYU professor who has made a career out of deflating AI hype, calling the Mythos capabilities "real but oversold." And David Sacks, the White House AI czar, who is reportedly pushing for Mythos to be fast-tracked to the Pentagon rather than delayed.

So who is right?

Honestly? Probably everyone, a little.

LeCun is right that the AI safety industry has an incentive to exaggerate threats. If your entire brand is built on being the "responsible" AI company, you need threats to be responsible about. Anthropic locked down Mythos in a way that generated more headlines than any product launch could have. The model that is too dangerous to release is the ultimate marketing hook.

But Amodei is also right that the capabilities are qualitatively different from what came before. Finding individual vulnerabilities is not new. Finding thousands of them autonomously, across different software stacks, in hours rather than months? That changes the math for every attacker on the planet. Even if smaller models can replicate some of this, Mythos apparently does it at a scale and speed that makes the comparison misleading.

And Marcus is right that both sides have blind spots. Anthropic overstates to justify its business model. LeCun understates because Meta has no comparable capability and dismissing Mythos serves his competitive interests.

My take: the debate itself is the story. We have reached the point where AI companies build things powerful enough to scare governments, and the leading minds in the field cannot agree on whether that fear is justified. This is not a healthy sign for the industry. It suggests we do not have the frameworks, the benchmarks, or the shared understanding needed to evaluate frontier AI capabilities.

If the smartest people building AI cannot agree on what it can do, the rest of us are flying blind. And that should concern everyone more than any single model.