Anthropic holds back risky model
Anthropic says it will not publicly release its latest cybersecurity‑oriented model because it could enable widespread hacking, citing risks like information leakage and misuse. That caution has shaken enterprise buying: the Pentagon’s distancing from Anthropic is driving defence customers toward smaller AI firms, and hyperscalers like AWS are publicly defending investments across multiple model vendors to preserve buyer optionality. The episode underlines that what vendors refuse to ship is now as important to platform procurement as what they do ship. (gizmodo.com) (theguardian.com) (reuters.com) (techcrunch.com)
Anthropic built a new model called Claude Mythos Preview, then decided not to release it to the public because the company says it can identify and exploit software flaws at a level that could spread hacking faster than defenders can respond. Anthropic said on April 7 that the model will stay limited to a small group of partners instead of going into broad release. (anthropic.com) The company says Mythos Preview can find and exploit zero-day vulnerabilities, which are software bugs unknown to the vendor, across every major operating system and every major web browser in its testing. Anthropic also said the model can turn known-but-unpatched flaws into working exploits, which is the step that turns a bug report into a break-in tool. (anthropic.com) So Anthropic created Project Glasswing, a restricted program that gives Mythos Preview to defenders instead of the general market. Anthropic says more than 40 organizations that build or maintain critical software infrastructure can use the model to scan and secure their own systems. (anthropic.com) Those launch partners include Amazon Web Services, Apple, Cisco, CrowdStrike, Google, Microsoft, Nvidia, Palo Alto Networks, JPMorganChase, and the Linux Foundation. Anthropic’s pitch is simple: if a model can act like a master lockpicker, you first hand it to the people who run the building. (anthropic.com) This would already be unusual on safety grounds alone, but it lands in the middle of Anthropic’s fight with the Pentagon. Reuters reported on April 9 that small defense artificial intelligence firms are now getting calls from generals, combatant commanders, and investors after the Pentagon’s relationship with Anthropic soured. (reuters.com) That Pentagon break matters because defense contracts are not just revenue; they are trust badges. Reuters reported that winning military work often helps companies sell to other United States government agencies and to commercial customers that want proof a model vendor can clear hard security reviews. (reuters.com) Now buyers are looking at a new question: not just which model is strongest, but which vendor will actually ship the strongest model to them. If Anthropic keeps its most cyber-capable system behind a fence, defense and enterprise customers may decide that a slightly weaker model they can fully deploy is more useful than a stronger one they cannot. (theguardian.com) Amazon Web Services is adjusting to that mood in public. TechCrunch reported on April 8 that Amazon Web Services chief executive Matt Garman defended Amazon’s $50 billion investment in OpenAI even though Amazon had already invested $8 billion in Anthropic, saying Amazon Web Services is used to partnering with companies it also competes with. (techcrunch.com) That is a cloud-computing version of not betting the whole house on one supplier. If one model company gets blocked by safety concerns, government politics, or product limits, Amazon Web Services can still tell customers it offers other models on the same platform. (techcrunch.com) Anthropic says its long-term goal is still to let users deploy Mythos-class models safely at scale, but it does not plan to make Mythos Preview generally available now. In 2026, that means the most important product decision in artificial intelligence may be the model a company refuses to sell. (anthropic.com)
