Anthropic announced Project Glasswing, a program that will provide selected companies with access to its in-development AI model, Claude Mythos Preview, tailored for cybersecurity uses.

Anthropic says rivals can also join Glasswing. It left open the possibility of OpenAI taking part.

On April 9, media outlet Sources, run by journalist Alex Heath (알렉스 히스), reported that Logan Graham said, "Glasswing can quickly transition into a third-party-led consortium in which all model providers participate." Graham leads Anthropic's Frontier Red Team.

Asked whether OpenAI could participate, he said he would welcome it "in principle" and added that OpenAI and Google DeepMind should take part, whatever form Glasswing evolves into. He stressed that companies already in competition with Anthropic are participating in Glasswing.

He said requests for access to Mythos have poured in from companies, institutions and infrastructure managers since the Glasswing announcement. Anthropic is also rapidly increasing staffing for the internal team that manages Glasswing. He said the response to Glasswing was better than expected.

Some have criticised Anthropic for providing a powerful model only to specific partners and excluding others. Graham rejected that, saying participants can use Mythos only for security work and cannot use it for product development or R&D. He added that Anthropic asked all participants to focus not on its interests but on the world's most important security challenges in the areas they represent.

Companies in Glasswing can use Mythos for security work, while Anthropic can use it for a range of purposes without restrictions. Graham acknowledged this imbalance of power but said it was standard practice at research companies with internal research capabilities.

Asked whether the reason for restricting access to Mythos was service costs, he said it was "not a decisive factor." He also stressed that Glasswing is not a profit-making business, and that employees joined it not for money but because it put into practice the mission of AI safety.

Anthropic did not say when it would release Mythos to the public. Graham said it must complete lengthy verification across three areas: model safety, verified use and monitoring of unauthorised use.