Imagine you are relying on the same software that helps run banks, hospitals, browsers, and cloud systems, only to learn that a new AI model can spot dangerous flaws in that code faster than almost everyone else. That is the practical concern around Anthropic’s unreleased Claude Mythos Preview. According to Anthropic, the model has found thousands of high-severity vulnerabilities, including flaws in every major operating system and web browser. The company says those capabilities matter because they could lower the cost and skill needed to find and exploit software weaknesses, making destructive cyberattacks easier to launch. Public sources also show why the issue has drawn attention: Anthropic says the fallout could reach economies, public safety, and national security.
The immediate audience is not ordinary chatbot users. Anthropic says Mythos Preview is being shared through Project Glasswing with named launch partners including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, along with more than 40 additional organizations that build or maintain critical software infrastructure. In simple terms, the model is being aimed at defenders first. The public rationale is that the same capability that could help attackers can also help trusted security teams find and fix flaws before others abuse them. Like a master key, a powerful tool can open the door for rescue or for theft; the danger depends on who holds it and what safeguards exist.
Where this fits is the deep plumbing of digital life: operating systems, browsers, cryptography libraries, web applications, and cloud infrastructure. Anthropic’s technical write-up says Mythos found a now-patched 27-year-old bug in OpenBSD, identified vulnerabilities in major browsers, and developed exploit chains on Linux and FreeBSD. Timing also matters. Anthropic presents Project Glasswing as an urgent response to what it describes as a rapid rise in frontier-model cyber capability. The company says AI models over the past year have become markedly better at reading code, spotting vulnerabilities, and developing exploits, and that Mythos represents a leap beyond earlier internal models.
How it works, in public descriptions, is more concrete than the broader warnings. Anthropic says it runs Mythos inside an isolated container with the target project and source code, asks it to find a vulnerability, and lets the model investigate, test hypotheses, and produce a bug report with proof-of-concept exploit steps. On the specific question of “how it escaped the sandbox,” public sources do not clearly confirm that the model itself escaped Anthropic’s testing environment. What Anthropic does publicly claim is narrower and still serious: one browser exploit written by Mythos chained four vulnerabilities and escaped both renderer and operating-system sandboxes, and another disclosed example combined a browser exploit, a sandbox escape, and local privilege escalation to let a malicious webpage write directly to the operating system kernel.
What comes next is cautious containment rather than broad release. Anthropic says Mythos Preview is unreleased, restricted to defensive work through Project Glasswing, and backed by up to $100 million in usage credits plus direct donations to open-source security groups. The company also says it plans to report publicly within 90 days on lessons learned and disclosed fixes where possible. For readers today, the clearest next step is practical rather than dramatic: treat patching, software updates, and supply-chain security as urgent housekeeping. Public sources support one conclusion with unusual clarity: tools for finding and exploiting flaws are improving quickly, and the safest response is faster defense, not passive waiting.