The competition over frontier artificial intelligence has moved beyond product rivalry into an open dispute over narrative control, safety authority, and who gets to define the boundaries of access.
OpenAI CEO Sam Altman has accused rival Anthropic of deliberately amplifying existential fears to market its newest model, Claude Mythos, while simultaneously restricting access to a tightly selected group of corporate partners.
Speaking on the “Core Memory” podcast hosted by Ashlee Vance, Altman characterized the messaging around Anthropic’s rollout as strategically alarmist.
“It is clearly incredible marketing to say, ‘We have built a bomb. We were about to drop it on your head. We will sell you a bomb shelter for $100 million to run across all your stuff, but only if we pick you as a customer,’” he said.
The framing points to a deeper ideological divide in Silicon Valley over whether frontier AI systems should be broadly distributed under controlled safeguards or concentrated within a limited set of vetted institutions.
Anthropic has opted for the latter approach with Claude Mythos. The company has withheld a public release, citing heightened cybersecurity capabilities within the model, particularly its ability to identify system vulnerabilities that could be misused. Instead, it introduced a restricted access framework known as Project Glasswing.
Under that programme, only 11 organizations were granted access, including Google, Microsoft, Amazon Web Services, Nvidia, and JPMorgan Chase. The selection spans cloud infrastructure, semiconductor manufacturing, and financial services, effectively placing frontier model access within a narrow layer of global digital infrastructure providers.
The rationale is rooted in the containment risk for Anthropic. As models become more capable of autonomous reasoning and system-level analysis, the potential for dual-use exploitation increases, particularly in cybersecurity contexts. Restricting access, in this view, becomes a precondition for controlled experimentation.
Altman rejects the implication that such restrictions are neutral or purely safety-driven. He argues that they also function as a form of narrative positioning that consolidates authority over AI deployment decisions.
“There are people in the world who, for a long time, have wanted to keep AI in the hands of a smaller group of people,” he said. “You could justify that in a lot of different ways, and some of it’s real like there are going to be legitimate safety concerns. But if what you want is like, ‘We need control of AI, just us, because we’re the trustworthy people, I think the fear-based marketing is probably the most effective way to justify that.”
The disagreement reflects a broader structural tension emerging in the AI sector: whether governance should be decentralized through broad access and iterative safeguards, or centralized through controlled deployment to a small set of institutions deemed capable of managing systemic risk.
OpenAI, where Altman leads operations, has generally pursued a more distributed model, releasing models to the public with layered safety constraints, usage monitoring, and incremental capability expansion. Even so, Altman acknowledged that not all systems would be broadly released.
“There will be very dangerous models that will have to be released in different ways,” he said. “The goal here is to benefit everybody and also to, I don’t say market this in a way but like get the world to come on this journey with us, and to say, ‘We are going to give you more powerful technology, there’s going to be responsibility that goes along with that.’”
“We are going to try to help set up the world for as much success as we can,” he added.
The contrast between the two approaches has sharpened as model capabilities accelerate. Anthropic’s Claude Mythos reportedly demonstrates heightened competence in identifying cybersecurity weaknesses, a capability that raises both defensive and offensive implications. In restricted-release environments such as Project Glasswing, those risks are managed through controlled exposure, limiting both the user base and the operational context.
Critics of restricted deployment models believe that they risk concentrating power in a small set of corporations, effectively turning frontier AI into an infrastructure layer governed by private gatekeepers rather than broadly accessible tools. Proponents counter that premature mass deployment could amplify misuse risks before sufficient containment mechanisms are established.
The rivalry has also become increasingly personal. Anthropic’s chief executive, Dario Amodei, previously held senior roles at OpenAI before founding the competing firm, embedding institutional memory and philosophical divergence into the competition itself.
Altman suggested that the broader discourse around AI risk has intensified tensions within the industry.
“I think the doomerism talk hasn’t helped. I think the way certain other labs talk about us hasn’t helped,” he said, adding, “I think the way Anthropic talks about OpenAI doesn’t help.”
Beyond corporate positioning, the dispute indicates an unresolved policy vacuum. Governments have yet to establish consistent global frameworks for frontier AI deployment, leaving major labs to effectively define their own governance regimes. In that environment, safety arguments, commercial strategy, and institutional trust become difficult to separate.
What is emerging is not just a product race, but a contest over legitimacy: who is authorized to build, release, and constrain systems that are increasingly embedded in critical infrastructure, financial networks, and cybersecurity operations. As capability thresholds continue to rise, the divide between open deployment and controlled access is likely to deepen, turning today’s rhetorical conflict into a defining fault line in the governance of advanced artificial intelligence.
