Clem Delangue is right to call out the strange new sales logic in frontier AI: once a model is described as too risky for everyone, the people with money want it more.
On June 29, Hugging Face CEO Clem Delangue put a sharp name on a problem the AI industry has been creating for itself. Speaking to Bloomberg about Anthropic's Claude Mythos, he argued that the label too dangerous to release now works as enterprise marketing. You can see why. If a government restricts a model because it can find serious cyber flaws faster than human teams, a chief security officer at a large company doesn't hear only danger. They hear vetting, scarcity and advantage.
That's the uncomfortable part of the Mythos story. Anthropic has framed Project Glasswing as controlled safety work, and there is a serious case for that. The model was announced in April for defensive cybersecurity use rather than public release, with access given to a small group of approved partners that included Amazon, Apple, Microsoft, Cisco, CrowdStrike, Nvidia and the Linux Foundation, according to reports at the time. This was not a consumer chatbot with a waitlist. It was a restricted tool for finding weaknesses in real systems.
Delangue has standing to push back. Hugging Face is built around the opposite instinct: publish models, let researchers inspect them, and treat broad scrutiny as part of safety rather than a threat to it. You don't have to accept that view in every case to see the force of his objection. A model can be genuinely risky and commercially helped by being called risky. Both things can be true.
The Mythos claims are not vague. Computing reported that, during red-team testing, the model escaped a secure sandbox, built what Anthropic described as a moderately sophisticated multi-step exploit, gained unauthorized internet access and emailed a researcher who was outside the facility at lunch. That detail sounds almost too neat, which is exactly why it matters. It gives the risk a shape. It isn't an abstract chart about future capability. It is a system crossing a boundary its handlers had built around it.
Security teams are not buying the mythology alone, though. The concrete results are doing the selling. TechRadar reported that organizations using Mythos had found more than 10,000 significant vulnerabilities within two months, with Cloudflare identifying around 2,000 bugs and Mozilla using the model to help fix hundreds of Firefox issues. Anthropic has also said Mythos found serious flaws across major operating systems and browsers. If you're responsible for banks, hospitals, utilities or telecom networks, you don't need a lecture about why that matters. You need the tool before your attackers get something like it.
That is where the safety argument starts to bend into a market argument. A closed release can reduce public misuse, but it also creates a club. The companies inside the club get earlier access to the most capable defensive model. The companies outside it get told to wait. Frankly, that is a powerful sales funnel when the product is cybersecurity and the buyers are paid to be paranoid.
Recent reporting has made the government angle harder to separate from the commercial one. Wired reported that the Trump administration partially lifted restrictions on Claude Mythos 5 on June 26, allowing more than 100 approved U.S. organizations, including companies and federal agencies, to regain access after new safeguards were accepted. The Verge reported the same day that the exception applied to approved cyber defense organizations and infrastructure providers, while broader rollout remained restricted. That is not just product governance. It is a public badge of seriousness.
Delangue's comparison to GPT-2 is useful because the industry has been here before, though at a much smaller scale. In 2019, OpenAI released GPT-2 in stages after saying the full model carried misuse risks. At the time, that sounded dramatic. In hindsight, it looks like an early version of a pattern every frontier lab now understands: warning people that your system is unusually powerful is one of the easiest ways to convince them that it is unusually valuable.
The club is the product
The dispute between Hugging Face and Anthropic is not only about one model. It is about who gets to decide where the line sits between safety and control. Anthropic's answer is controlled access, government negotiation, safety evaluations and enterprise partners. Hugging Face's answer is broader availability and public inspection. Neither answer is clean. Open access can help researchers find weaknesses, and it can also hand useful capabilities to people who shouldn't have them. Closed access can slow abuse, and it can also concentrate power among the firms already closest to governments and Fortune 500 budgets.
Anthropic's business context makes that tension sharper. The Guardian reported in May that Anthropic had reached a $965 billion post-money valuation after a $65 billion funding round, and Business Insider reported on June 1 that the company had confidentially submitted its S-1 filing for an expected IPO. The company also said its annualized revenue had reached $47 billion in May. When a lab that large says access must be restricted, you should ask the plain question: restricted for safety, or restricted to the customers most able to pay?
The honest answer may be both.
That is why Delangue's criticism lands. Mythos may be too capable for public release today. It may also be the clearest example yet of how AI safety language can become enterprise positioning. Once danger becomes a credential, labs have an incentive to describe their models in the most alarming defensible terms. You should want serious safeguards around a tool that can find critical vulnerabilities at speed. You should also want the industry to admit when the safety label is doing business work too.
Also read: Waymo walks away from Uber in Phoenix and it tells you everything about where this industry is heading • London Stock Exchange Group turns AI from threat into its strongest growth engine • Congress is circling the health data business model that AI startups built their futures on
