Anthropic said on April 7 that it would not release its latest artificial intelligence model, Mythos, arguing that the system is too dangerous to be made public and could be misused in ways that threaten economies, public safety and national security. Instead, the company said it is opening the model only to a small group of organizations through its Project Glasswing initiative for cybersecurity testing.
The move puts Anthropic in unusual territory for a major AI developer. The company released a 245-page technical document with the announcement and described Mythos as operating like a senior software engineer. It also said the model scored 31 percentage points higher than Opus 4.6 on the USAMO 2026 Mathematical Olympiad and can outstrip all but the most skilled humans at identifying and exploiting software vulnerabilities.
Anthropic said Mythos found critical faults in every widely used operating system and web browser it tested, and that 99 percent of those vulnerabilities have not yet been patched. An assessment by the U.K.'s AI Security Institute found the model succeeded in expert-level hacking tasks 73 percent of the time, a level no AI model had reached before April 2025. The company says the first Project Glasswing group includes Microsoft, Google, Apple, Amazon Web Services, JPMorgan Chase and Nvidia.
The decision to withhold Mythos echoes OpenAI's temporary withholding of GPT-2 in 2019, but cybersecurity experts remain split over whether this is a sharp break from earlier AI releases or the next step in a trend they have already been warning about. Anthropic says the limited access is meant to let selected organizations scan for weaknesses and patch them before flaws become public, a defensive use case that contrasts with the model's own apparent offensive strength.
That tension is already shaping the reaction. On Thursday after the announcement, German banks said they were consulting authorities and cyber experts about the risk. For Anthropic, Mythos is now both a warning and a test: whether powerful AI can be held back long enough to reduce harm, or whether keeping it behind the curtain only underscores how quickly the technology has outgrown the old release playbook.
