Anthropic's Mythos: locked down

# Anthropic’s Mythos: locked down Anthropic has done something unusual in the race to build more powerful artificial intelligence systems: it announced a new flagship model and then declined to release it to the public. On April 7, 2026, the company said its new model, Claude Mythos Preview, is strong enough at finding software flaws that broad access could help criminals and state-backed hackers as much as defenders. Instead, Anthropic is putting the model inside a restricted cybersecurity program called Project Glasswing. (anthropic.com) That decision tells you where the frontier of artificial intelligence is moving. For years, the main public demos of large language models were writing, search, and coding help. Mythos points at a different use case: reading huge codebases like a tireless security auditor, spotting the tiny mistakes that let attackers break in. Anthropic says that capability is now strong enough that the company wants defenders to get a head start before the wider market does. (cnbc.com, bloomberg.com) To understand why Anthropic is locking Mythos down, it helps to start with what software security actually looks like. Modern software is built from millions of lines of code, plus outside components pulled in from open-source projects and vendor libraries. A single overlooked mistake can sit quietly for years, like a bad lock hidden inside the wall of a house, until someone figures out how to use it. (anthropic.com) Security teams spend much of their time looking for those hidden mistakes before attackers do. They use scanners, tests, code review, and human researchers to hunt for weaknesses such as memory errors, broken permissions, and unsafe ways of handling data. The problem is scale: the amount of code that matters to banks, cloud providers, operating systems, browsers, and industrial systems is far larger than any human team can inspect line by line. (anthropic.com) That is where advanced artificial intelligence changes the picture. A strong coding model can read source code, reason across files, generate test cases, and try attack paths much faster than a person. If it is good enough, it stops being just an assistant for programmers and starts acting more like an always-on junior security researcher that can check thousands of places in parallel. (techcrunch.com, anthropic.com) The same capability cuts both ways. A model that helps defenders find weak spots can also help attackers find them first, explain how to exploit them, and automate parts of the break-in process. Anthropic’s position is that Mythos has crossed far enough into that territory that a normal public launch would be reckless, at least for now. (cnbc.com, bloomberg.com) Anthropic describes Mythos as its most capable frontier model so far, and its system card says the model shows a large jump over Claude Opus 4.6, the company’s previous top tier. In Anthropic’s framing, Mythos is still a general-purpose model, not a tool built only for cybersecurity, but its coding and reasoning abilities make it unusually effective at security work. (anthropic.com, techcrunch.com) The company’s public claims are striking. Anthropic says Mythos has identified thousands of zero-day vulnerabilities, including critical ones, across major operating systems and web browsers. A zero-day vulnerability is a flaw that defenders do not know about yet, which means there is no patch waiting if an attacker finds it first. (techcrunch.com, thehackernews.com) Anthropic and Bloomberg both gave concrete examples meant to show the age and subtlety of the bugs Mythos can surface. According to Bloomberg’s reporting, Anthropic said Mythos found a 27-year-old bug in critical internet software and a 16-year-old vulnerability in game software code that automated testing had scanned five million times without catching. Those examples are the kind of detail security engineers pay attention to, because old bugs in mature code are usually the hardest to shake loose. (bloomberg.com) Instead of a public rollout, Anthropic created Project Glasswing as a gated testing and deployment channel. The company says the program is aimed at securing “the world’s most critical software” and is giving early access to organizations that build or maintain infrastructure used by billions of people. Anthropic is also committing up to $100 million in usage credits and $4 million in donations to open-source security organizations as part of the effort. (anthropic.com) The launch partners show the kind of software Anthropic wants Mythos pointed at first. Anthropic lists Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks among the initial partners. It also says more than 40 additional organizations that build or maintain critical software infrastructure have been given access. (anthropic.com, techcrunch.com) That coalition matters because the security problem is shared. Cloud companies run the computers other businesses depend on. Operating system makers and browser makers sit underneath everyday internet use. Open-source maintainers keep foundational code running across the web. Security vendors see attacks across thousands of customers. If Mythos is going to be tested safely, these are the places where it can find flaws with the biggest downstream effect. (anthropic.com) Anthropic says partners will use Mythos for defensive security work, including scanning both first-party and open-source systems. The company also says it plans to share lessons from the program with the broader industry, which suggests Project Glasswing is meant to be both a live defense effort and a way to learn what safeguards are needed before any wider deployment. (techcrunch.com, anthropic.com) CNBC reported that Anthropic’s executives debated the release internally before deciding on the restricted approach. Dianne Penn, Anthropic’s head of research product management, told CNBC there was “a lot of internal deliberation,” and described the program as a first step to give defenders a head start. That is consistent with Bloomberg’s report that Anthropic does not currently plan to make Mythos generally available and wants feedback from Glasswing to shape guardrails first. (cnbc.com, bloomberg.com) The timing also matters because Mythos was