AI rollouts get held back

Artificial intelligence labs usually race to ship first, but in April 2026 two of the biggest labs did the opposite and slowed down because their own models looked too useful for hacking. Axios reported on April 9 that OpenAI is planning a phased rollout of a new model over cybersecurity worries. (axios.com) Anthropic made the slowdown public first. On April 7, it said its new model, Claude Mythos Preview, would go only to a limited set of partners instead of a broad public release. (anthropic.com) (cnbc.com) The reason was not that the model wrote prettier emails or faster summaries. Anthropic said Claude Mythos Preview had already found thousands of major vulnerabilities in operating systems, web browsers, and other software. (anthropic.com) (reuters.com) That creates a simple problem: the same system that helps a defender find a hole can help an attacker find the same hole first. Anthropic’s answer was a program called Project Glasswing that gives early access to defenders before a wider launch. (anthropic.com) (cnbc.com) The first users are not random subscribers. Anthropic said launch partners include Amazon Web Services, Apple, Google, Microsoft, Nvidia, CrowdStrike, Palo Alto Networks, JPMorganChase, Cisco, Broadcom, and the Linux Foundation. (anthropic.com) Anthropic also said it extended access to more than 40 additional organizations that build or maintain critical software infrastructure. It paired that with up to $100 million in usage credits and $4 million in donations to open-source security groups. (anthropic.com) (reuters.com) This is not an improvised policy change. Anthropic updated its Responsible Scaling Policy on February 24, 2026, and said stronger safeguards should kick in when models cross higher capability thresholds. (anthropic.com) OpenAI has a similar playbook on paper. In its updated Preparedness Framework from April 15, 2025, OpenAI said it tracks frontier-model risks in three standing categories, including cybersecurity, and ties deployment to evaluations and safeguards. (openai.com) The backdrop is that labs now think cyber capability is moving from “helpful coding assistant” to “machine that can scan the internet for weak locks.” Axios reported on March 29 that top artificial intelligence and government officials were warning that new models from Anthropic, OpenAI, and others were becoming frighteningly good at hacking sophisticated systems at scale. (axios.com) Anthropic’s own rollout shows what “slow down” means in practice. The company said its eventual goal is still to let users deploy Mythos-class models at scale, but only after defenders get a head start and the model is hardened against misuse. (reuters.com) (anthropic.com) So the new pattern in frontier artificial intelligence is not just bigger launches and louder demos. It is labs treating some models like powerful tools in a locked cabinet: useful enough to matter, and dangerous enough that handing them to everyone at once no longer looks responsible. (axios.com) (openai.com)