Anthropic’s glasswing expansion: 10,000 vulnerabilities and the AI cyber-threat countdown
Anthropic just dropped a number that should make every IT department on the planet break into a cold sweat: 10,000.
That’s how many "high- or critical-severity" security flaws the company’s initial 50 partners found after scanning their codebases with Claude Mythos Preview. Looking at this practically, we aren't just talking about a few leaky buckets; we’re talking about a structural rot in the world’s most important software that we are only just beginning to map out.
In the spirit of clarity, Anthropic is now tripling down. They are expanding "Project Glasswing"—their collaborative effort to secure global infrastructure—to approximately 150 new organizations across 15 countries. We’re moving beyond tech firms and into the scary stuff: power, water, healthcare, and communications.
The "Glasswing" initiative relies on Claude Mythos Preview, a model so potent for cyber-defense (and, by extension, cyber-offense) that Anthropic won't even let the general public touch it. It’s the high-security VIP room of AI models, reserved for those who can prove they won't use it to accidentally—or intentionally—topple a government.
Here’s the challenge we need to navigate: Anthropic is openly admitting that "Mythos-class" models from competitors are likely only 6 to 12 months away. The kicker? They expect those models will be released without the same safety "glass" that Anthropic has spent millions building.
It feels like a high-stakes game of beat-the-clock. If the good guys don't use AI to patch the holes today, the bad guys will use AI to exploit them tomorrow. In the current landscape, we should prepare for a world where the very definition of "secure code" changes every few months as models get smarter.
While saving the Western world’s power grid, Anthropic is also aggressively planting flags in the East. They just tapped KiYoung Choi, a veteran of Snowflake, Google Cloud, and Microsoft, to lead their new Seoul office.
Koreans are apparently using Claude at 3.5 times the expected rate for their population. That’s not just casual chatting; it’s heavy-duty technical and creative work. From Law&Company’s legal assistants to SK Telecom’s customer service bots, Anthropic is clearly aiming for a total enterprise takeover in one of the world's most sophisticated AI markets.
Ultimately, Anthropic is trying to play both the neighborhood watch and the global real estate mogul. They’re warning us that the internet is a house of cards while simultaneously selling us the glue to hold it together.
It's a noble goal, but as we've seen in every tech arms race, the "permanent advantage for defenders" usually only lasts until the next update.
Sources: Expanding Project Glasswing, Anthropic appoints KiYoung Choi as Representative Director of Korea.
