Anthropic Holds Back Mythos
Anthropic says it is not broadly releasing its new Claude Mythos model because the company believes the model can autonomously find and exploit software vulnerabilities at an unusually high level. That decision prompted Anthropic to form Project Glasswing and limit access to a cybersecurity consortium rather than open distribution, signalling a rare retreat from full public release. The move matters because it shows frontier code-generation models are now being treated as a security risk rather than just productivity tools, changing how engineers and vendors think about access and oversight. (qz.com) (cnet.com)
Anthropic built a new Claude model, then decided not to put it on the open market. On April 7, the company said Claude Mythos Preview would go only to a closed group of security partners because it can find and exploit software flaws at a level Anthropic says exceeds nearly all human hackers. (anthropic.com) Project Glasswing is the fence Anthropic built around that model. The launch group includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks, plus more than 40 other organizations that maintain critical software infrastructure. (anthropic.com) The basic problem is simple: every major program has bugs, and a small share of those bugs can become break-in points. Anthropic says Mythos Preview has already found thousands of high-severity vulnerabilities, including flaws in every major operating system and web browser. (anthropic.com) A software vulnerability is like a hidden side door in a bank that even the bank forgot was there. Anthropic says Mythos does not just spot those doors in code reviews; it can also work out how to open them, which is why the company is treating release as a security decision instead of a product launch. (nbcnews.com) Anthropic says the model’s cyber strength comes from stronger coding and reasoning, not from being a special-purpose hacking bot. In its system card, published April 7, the company called Mythos Preview its most capable frontier model to date and said the jump in capability was large enough to cancel a general release. (reason.com) This is not Anthropic improvising from scratch. On February 24, the company updated its Responsible Scaling Policy, a rulebook that says stronger models trigger stronger safeguards as capabilities rise from chat answers to web use, code execution, and autonomous multi-step actions. (anthropic.com) That policy uses “Artificial Intelligence Safety Levels,” which are internal tiers for how tightly a model must be handled. Anthropic’s latest policy says higher-capability systems need stricter controls on deployment and on theft or misuse, which helps explain why Mythos landed in a consortium instead of a public application programming interface. (anthropic.com) The story also broke in public before the official launch. On March 26, Fortune reported that Anthropic had left draft material about Mythos in an unsecured, publicly searchable data store, and Anthropic later confirmed the unreleased model existed. (fortune.com) When Anthropic finally announced the plan on April 7, it attached money to the containment strategy. The company said it would provide up to $100 million in Mythos usage credits and another $4 million in direct donations to open-source security organizations so partners can scan and harden widely used code. (anthropic.com) Anthropic is pitching this as a race between defenders and attackers. Chief executive Dario Amodei said the upside is a more secure internet, but CNBC reported the company spent time debating the release internally because a broad rollout could hand the same capability to criminals. (cnbc.com) The unusual part is not that an artificial intelligence company built a stronger coding model. The unusual part is that one of the biggest model makers looked at its own product and decided the safest launch was something closer to controlled access for critical infrastructure than a normal software release. (anthropic.com)
