Anthropic built its most capable AI model to date and chose not to sell it.
The company announced Claude Mythos alongside Project Glasswing, a controlled program giving select organizations early access to Mythos Preview. The goal is to let defenders find and fix critical software vulnerabilities before bad actors get access to comparable capabilities. Partners include AWS, Apple, Cisco, Google, JPMorgan Chase, Microsoft, NVIDIA, and Palo Alto Networks.
The testing results are the part that got the industry's attention. In just a few weeks, Mythos Preview autonomously identified thousands of previously unknown zero-day vulnerabilities across every major operating system and every major web browser. One of those was a 27-year-old flaw in OpenBSD that would have allowed any attacker to remotely crash machines running it, simply by connecting to the device.
When Anthropic says autonomously, they mean it. In one documented test, engineers pointed Mythos at a codebase before going home for the day. By morning, the model had produced a fully functional remote code execution exploit with no human involvement at any stage after the initial request.
Mythos also demonstrated something security professionals call vulnerability chaining, connecting a series of individually minor software flaws into a single devastating attack path. IBM's Dave McGinnis, VP of Global Managed Security Services, called it a genuine step change: "It's not like they created the bugs. The people who wrote that code didn't know those things were there."
The reason Anthropic is withholding public access is straightforward. The same reasoning power that makes Mythos exceptional for defense makes it dangerous in the wrong hands. Anthropic's logic: give defenders early access to patch the most critical vulnerabilities before Mythos-class capabilities reach a broader audience.
Industry analysts at Forrester said Anthropic is now "the most important partner for every cybersecurity company." Others were less generous. AI critic Ed Zitron called the announcement marketing-dressed alarm. Both reactions can be true at the same time.
IBM's analysts estimate other frontier labs are months away from comparable capabilities. The window for defenders to get ahead is narrow.
For enterprise security teams, this is a 2026 planning problem. The time between a vulnerability's public discovery and its first observed exploit dropped from 771 days in 2018 to single-digit hours by 2024. Models like Mythos compress that window further. Traditional patching cycles measured in days or weeks are already struggling to keep pace.
The question worth asking inside your organization is not whether AI will change the cybersecurity landscape. It already has. The question is whether your security posture is built for the speed this environment now requires.
Anthropic committed $4 million to open-source security foundations through the Glasswing initiative. The company says its goal is eventually enabling Mythos-class capabilities in future public models, once appropriate safeguards are in place. If external forecasts are anywhere close, a public-facing Mythos-class model could arrive before the end of 2026. That is not much runway for organizations still running legacy vulnerability management programs.
