Anthropic on Tuesday formally announced Mythos Preview and opened it first to an industry consortium called Project Glasswing, a group of more than 40 companies and institutions that will get private access before any general release.
The partners include Microsoft, Apple, Google, Amazon Web Services, the Linux Foundation, Cisco, Nvidia and Broadcom. Anthropic said the staggered rollout is meant to give developers time to turn Mythos Preview loose on their own systems and find weaknesses, exploit chains and other attack paths before the model reaches wider use.
The company is treating the release as a controlled test of a system it believes marks a real step up. Dario Amodei said Mythos Preview is a particularly big jump, and said Anthropic trained it to be strong at code, with cyber ability emerging as a side effect. Logan Graham said the model can discover vulnerabilities, build attack chains and proofs of concept, develop exploits, carry out penetration testing, assess endpoint security, hunt for misconfigurations and evaluate software binaries without source code.
Graham said Anthropic wants to prepare for a world where those capabilities are broadly available in 6, 12 or 24 months. He said the point is not the model or Anthropic itself, but the shift underway in security. “Many things would be different about security,” he said in substance, warning that some of the assumptions behind modern security systems could break. He also said Mythos Preview has already done things that a senior security researcher would be able to accomplish.
Read Also: Anthropic Confirms Is Claude Down With Around 3,000 Downdetector Reports
That framing matters because Anthropic is not just releasing another AI tool. It is drawing on the logic of coordinated vulnerability disclosure, using Project Glasswing as a way to let defenders study the system before it is placed in the broader market. The company said that more powerful models are coming from it and from competitors, and that the industry needs a plan to respond.
The release follows leaked revelations at the end of March that Anthropic had developed a powerful new Claude model. Anthropic said models from multiple companies have increasingly been able to find flaws in code and suggest fixes or exploitation strategies, and framed the collaboration as part of a larger response to how advancing AI is changing software security and digital defense. Graham said that if the process is not handled carefully, it could become a meaningful accelerant for attackers.
For now, the work starts behind closed doors. Project Glasswing gives Anthropic a way to test how dangerous Mythos Preview could become in the hands of defenders and attackers alike, before the model is widely available.
