Anthropic is widening access to Claude Mythos, but only through a gate. That may be the clearest sign yet that frontier AI security tools are becoming enterprise infrastructure, not public software.
Anthropic has expanded Project Glasswing to roughly 150 additional organizations across more than 15 countries, moving Claude Mythos Preview from a tightly watched experiment into a broader test of how powerful cyber AI should be distributed. The important detail is not just who gets access. It is who does not.
Claude Mythos is not being released like a normal model. It is being handed to selected companies, public agencies, and infrastructure providers that meet Anthropic's security requirements. That matters because the model is designed to find and exploit software vulnerabilities, which makes it useful for defenders and dangerous in the wrong hands. In cybersecurity, the same skill can patch a system or break into it.
As Anthropic said in its June 2 update, the new group covers sectors that were not well represented in the first wave, including power, water, healthcare, communications, and hardware. That is a practical admission that the risk is no longer limited to big software vendors. The code that keeps hospitals, utilities, telecom networks, banks, and chip companies running is now part of the AI security race.
The first Project Glasswing group included roughly 50 partners. Anthropic says those partners have already used Claude Mythos Preview to find more than 10,000 high or critical severity security flaws. That figure is striking, but the deeper issue is capacity. Finding vulnerabilities at that pace is not the same as fixing them. Every bug still has to be verified, disclosed, prioritized, patched, and deployed without tipping off attackers.
That turns Mythos into more than a technical achievement. It turns it into an operating model. Anthropic is effectively saying that the most capable cyber models should be placed behind a trust layer, where access depends on purpose, sector, and security controls. That is very different from the standard software playbook, where broader access usually means faster adoption and bigger network effects.
There is a business logic here as well. Anthropic has committed $100 million in model usage credits to Project Glasswing and additional participants, but it has also said Mythos Preview will later be available to participants at $25 per million input tokens and $125 per million output tokens. This is not just a safety program. It is the outline of a restricted, high-value enterprise security channel.
For large infrastructure operators, that could be an easy budget conversation. A serious software flaw can cost far more than model usage. For Anthropic, the advantage is just as clear: it gets real-world security data, close relationships with critical institutions, and a reason to sell access to models that are too sensitive for ordinary public release.
SK Telecom shows why the program is spreading
SK Telecom is one of the more useful examples of why this model is gaining ground. The South Korean carrier confirmed it has joined Project Glasswing and gained access to Claude Mythos Preview to help identify vulnerabilities in its systems before they are exploited. Light Reading reported that SKT is doing the work under the project's governance protocols, which is exactly the kind of controlled deployment Anthropic wants to highlight.
The timing is not accidental. SK Telecom has been rebuilding its security posture after a major USIM data breach that affected about 23 million subscribers. The company has allocated about 700 billion won, roughly $453 million, over five years to strengthen cybersecurity, including AI-based monitoring and a move toward a Zero Trust framework. In that context, Mythos is not a shiny experiment. It is part of a defensive rebuild.
South Korea's participation also widens the strategic picture. Reports have named Samsung Electronics, SK Hynix, SK Telecom, and Korea's KISA among the Korean organizations connected to the expanded effort. That brings semiconductors, telecoms, and public cyber response into the same controlled AI security program. For a country whose chip and network infrastructure sits near the center of global supply chains, that is not a small development.
The competitive question is whether this creates a moat for Anthropic. It probably does, at least for now. If the company becomes the trusted supplier of frontier cyber models to governments, banks, telecoms, cloud providers, chipmakers, and infrastructure operators, it is building relationships that are hard to copy with a cheaper chatbot or a slightly faster coding model.
But the same structure could become a liability. A gated system raises hard questions about who gets protected first, who is left waiting, and how much critical security knowledge ends up concentrated inside one private AI company. Anthropic says other companies may have Mythos-class models within 6 to 12 months, and some may release them without strong safeguards. If that happens, controlled access will look less like caution and more like the first stage of a much larger scramble.
The next thing to watch is not whether Mythos finds more bugs. It almost certainly will. The question is whether the institutions receiving access can fix software fast enough to make the defensive advantage real. If they can, Project Glasswing becomes a model for frontier AI deployment. If they cannot, the industry will have proved that finding the weakness was the easy part.
Also read: Jensen Huang says the AI selloff has opened a buying window • AI fears are choking the software buyout machine • Microsoft is making OpenAI optional inside its AI stack
