Anthropic held back a model
Social reporting says Anthropic withheld a powerful model and restricted access to about 40 partner companies after a leak and safety concerns, with officials urging urgency and some public criticism of the decision. (x.com) The conversation included both alarm about the leak and pushback calling the move an overreaction, reflecting divided industry responses. (x.com)
Anthropic has held back its newest artificial intelligence model from a public launch and limited access to a small security group instead. (anthropic.com) The company said on April 7, 2026 that Claude Mythos Preview is its “most capable frontier model to date” and that the jump in capability was large enough that it would not make the model generally available. (anthropic.com) Anthropic put the model inside Project Glasswing, a defensive cybersecurity program with 12 named launch partners including Amazon Web Services, Apple, Google, JPMorganChase, Microsoft, Nvidia and Palo Alto Networks. It also said it extended access to more than 40 additional organizations that build or maintain critical software infrastructure. (anthropic.com) The reason for the restriction is straightforward: Anthropic says the model is unusually good at finding serious software flaws, the kind attackers can use to break into systems before patches exist. The company said those capabilities had already helped identify and fix vulnerabilities across hardware and software. (anthropic.com) This story started in public two weeks earlier, when Fortune reported that draft Anthropic materials had been left in a publicly searchable data cache. Anthropic said a “human error” in its content management system configuration exposed the drafts. (finance.yahoo.com) Those leaked drafts described Mythos as a “step change” beyond Anthropic’s earlier Claude models and said the company believed it posed major cybersecurity risks. Anthropic later confirmed the model was being tested with early-access customers before the April 7 announcement. (finance.yahoo.com) Anthropic’s public rollout sharpened the split in industry reaction. Dianne Penn, Anthropic’s head of research product management, told CNBC there was “a lot of internal deliberation,” while critics on social media argued the company was overstating the danger or turning a leak into a marketing event. (cnbc.com) Anthropic framed the decision as part of a broader push to keep advanced models in narrower channels when they cross internal safety thresholds. The system card says Mythos was evaluated not just for coding and cyber ability, but also under the company’s Responsible Scaling Policy and alignment testing. (anthropic.com) Project Glasswing also comes with money attached: Anthropic said it is committing up to $100 million in usage credits and $4 million in donations to open-source security groups. The company said it will share lessons from the restricted rollout with the wider industry. (anthropic.com) For now, the practical change is simple: Anthropic has a stronger model, but most users cannot touch it. The company is treating Mythos as a tool for selected defenders first, not a product for the general market. (anthropic.com)
