BBC questions Anthropic's safety stance

Anthropic is the AI company that has spent years telling the world it takes safety more seriously than everyone else. That pitch just ran into its hardest test. On May 1, BBC News aired a segment asking whether Anthropic is “the new Dr Frankenstein,” after the company revealed a powerful model called Claude Mythos and then said it was too risky for general release. (youtube.com) ### What actually triggered this? The immediate trigger was Claude Mythos Preview, announced on April 7. Anthropic described it as its most capable frontier model yet, with a big jump over its previous top system. But instead of shipping it broadly, the company said it would keep access tight and use it only with selected partners in a defensive cybersecurity program. (www-cdn.anthropic.com)pdf)) ### Why is Mythos different from a normal chatbot? Because this is not mainly a “write me an email” model. Anthropic says Mythos is unusually strong at finding and fixing software vulnerabilities. The BBC segment centers on the claim that, during testing, the system uncovered thousands of hidden weaknesses in widely used software and could outperform humans at som(www-cdn.anthropic.com)at mostly summarizes documents and answers questions. (youtube.com) ### So why the Frankenstein framing? Basically, because Anthropic is arguing two things at once. First: this model is useful enough that major infrastructure and security players should start using it now. Second: this same model is dangerous enough that the public should not get it. That tension is exactly what makes the story sticky. If your brand is “we are the careful lab,” people immediately ask the(youtube.com)anthropic.com) ### Isn’t holding it back the safe move? Yes — but only partly. Not releasing a dangerous model widely is a real safety action. Anthropic’s own system card says the capability jump led it to decide against general availability. But the catch is that this does not remove the risk. It changes who gets access, under what controls, and for what purpose. The company is still developing the model, still test(anthropic.com)rs. (www-cdn.anthropic.com) ### Who does get access? Anthropic put Mythos inside Project Glasswing, a new security initiative with launch partners including Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, and the Linux Foundation. It also says more than 40 additional organizations that build or maintain critical software infrastru(www-cdn.anthropic.com)ource security groups. (anthropic.com) ### Why does that still make people uneasy? Because “limited release” is not the same thing as “contained.” The more capable a system becomes at offensive-style cyber tasks, the more the public starts worrying about leakage, misuse, theft, or simple loss of control. Anthropic’s whole case is that the best defense is to get defenders a head start. But critics hear something else — that a private company(anthropic.com)g cyber advantage before governments and the public have settled the rules. That last part is an inference from Anthropic’s rollout and the BBC framing, but it is the core political issue here. (anthropic.com) ### What does this say about Anthropic’s safety stance? Turns out the story is less “Anthropic abandoned safety” than “Anthropic’s safety model is being stress-tested in public.” The company still has a Responsible Scaling Policy, and it updated that policy again on April 29 to formalize more external review and board briefings. But Mythos shows the harder truth: a safety-first lab does not avoid dange(anthropic.com)to convince everyone its guardrails are real. (anthropic.com) ### Bottom line? The BBC segment lands because it hits the contradiction people already feel about frontier AI. Anthropic wants credit for stopping short. But once a company says, in effect, “we built something too risky for public release,” the argument is no longer about branding. It is about whether private labs should be the ones deciding how close to the edge is acceptable.