Anthropic Faces Skepticism Over Decision to Withhold Powerful Mythos AI Model Due to Security Concerns

San Francisco-based AI company Anthropic said this week it has built a powerful AI model called Mythos, but will not make it publicly available, citing cybersecurity risks. The decision quickly drew public and political attention, including engagement from U.S. Treasury Secretary Scott Bessent and Reform UK MP Danny Kruger, as the company’s claims circulated widely.

 

Some AI critics and scientists questioned whether the announcement was primarily a safety-driven move or a bid to strengthen the company’s profile with investors. Among them, AI expert Gary Marcus argued that the messaging resembles a familiar pattern in the sector, comparing Anthropic’s approach to that of rival OpenAI and suggesting that hype and exaggeration can shape public perception.

 

Dr. Heidy Khlaaf, chief AI scientist at the AI Now Institute, said Anthropic’s marketing post relied on “purposely vague language that obscures evidence,” which she said makes it harder to evaluate the stated rationale for withholding Mythos. The criticism centers on the gap between the seriousness of the security justification and the limited detail provided publicly about what Mythos can do and what specific risks it poses.

 

Questions about the company’s security posture have also been sharpened by a recent incident involving its flagship model family. In early April, Anthropic accidentally leaked part of Claude’s internal source code. The company said the leak did not expose sensitive customer data, but the episode added to scrutiny as Anthropic asks the public to accept its judgment that Mythos is too risky to release.

 

On the technical claims, Jameison O’Reilly, described as an offensive cybersecurity expert, said Mythos is a genuine development while disputing the weight of some of the company’s assertions. He questioned the significance of claims such as identifying thousands of “zero-day vulnerabilities,” and said such vulnerabilities are rarely critical to achieving objectives in real-world cybersecurity engagements.

 

Separately, Anthropic is also navigating constraints that could shape product decisions. The company has faced limited computing capacity and has implemented usage caps for its Claude models, which may also factor into why Mythos is being held back. The combination of security framing, external skepticism, and operational limits has turned the Mythos announcement into a broader test of credibility for a major AI developer operating under intense political and market attention.