Frontier models gated access
Anthropic said its new Mythos model is too risky for public release and will be distributed via a restricted preview, and OpenAI is reportedly planning a cybersecurity-focused model with a staggered, partner-only rollout. That pattern — powerful models arriving through vetted enterprise channels rather than open release — is starting to shape expectations around who gets access and under what controls. ( )
Anthropic did something the artificial intelligence industry hates doing in public: it built a stronger model and then refused to broadly release it. On April 7, 2026, the company said Claude Mythos Preview would go only to selected organizations through a program called Project Glasswing. (anthropic.com) The reason was not bad chatbot answers or copyright fights. Anthropic said Mythos was unusually strong at finding software flaws and turning them into working attacks, including zero-day vulnerabilities, which means bugs nobody had publicly found yet. (anthropic.com) In its own testing, Anthropic said Mythos could identify and exploit zero-day vulnerabilities in every major operating system and every major web browser. The company said some of the bugs were 10 to 20 years old, and the oldest patched example it cited was a 27-year-old bug in OpenBSD. (anthropic.com) Anthropic did not send this model to the usual mass market of developers and hobbyists. NBC News reported that the first access was going to companies including Microsoft, Nvidia, and Cisco, and that Project Glasswing would give more than 50 organizations over $100 million in usage credits. (nbcnews.com) That is a sharp break from the last three years of artificial intelligence launches, when the normal pattern was a public application programming interface first and tighter controls later. Anthropic’s own April 7 post described Mythos as a “general-purpose language model,” but the rollout treated it more like a sensitive security tool than a normal product launch. (anthropic.com) The policy backdrop matters here. In February 2026, Anthropic updated its Responsible Scaling Policy, which says stronger safeguards should kick in when models cross specific capability thresholds instead of waiting for a crisis after release. (anthropic.com) OpenAI has been moving in the same direction on paper. In its April 15, 2025 Preparedness Framework update, OpenAI said cybersecurity is one of the “Tracked Categories” where it measures frontier-model capability and builds safeguards before deployment. (openai.com) Now that paper is starting to look like product strategy. Axios reported on April 9, 2026 that OpenAI is finalizing a cybersecurity product with advanced capabilities and plans to release it only to a small set of partners. (axios.com) Put those two moves together and a new access model is coming into view. The most capable systems may reach banks, cloud providers, chip companies, and government-linked partners first, while the public gets weaker versions later or not at all. (axios.com) That changes the old argument over “open versus closed” artificial intelligence into a newer one about clearance. The practical question is no longer only who can build the strongest model, but who gets vetted, monitored, logged, rate-limited, and trusted to use it. (anthropic.com) If this pattern holds, frontier artificial intelligence will start to look less like an app store and more like a restricted lab. Anthropic’s April 2026 Mythos launch is the clearest sign yet that, for some models, the product is not just the model itself but the gate around it. (anthropic.com)
