Anthropic expands Project Glasswing to protect critical infrastructure

Anthropic is urgently expanding its defensive shield before AI-powered cyberweapons hit the open market.

Back in early April, the AI lab kicked off Project Glasswing, an initiative that handed access to its unreleased Claude Mythos Preview model to about 50 initial partners. In just a few weeks, that small group deployed the model across their codebases and surfaced more than 10,000 high- or critical-severity flaws.

Now, Anthropic is expanding the program to roughly 150 new organizations across more than 15 countries. The new roster intentionally targets critical infrastructure sectors that were largely missing from round one, including power grids, water utilities, healthcare systems, communications, hardware makers, and open-source software maintainers.

These aren't casual web apps. Anthropic estimates that a single breach in most of these codebases could impact over 100 million people.

The 6-to-12-month warning

The rush comes down to a grim deadline. Anthropic warns that within 6 to 12 months, other AI labs will build Mythos-class models, and many may release them without strict safety guardrails against malicious use.

Basically, offensive actors are about to get cheap, lightning-fast tools capable of tearing apart software at machine speed. If defenders don't harden critical codebases before that happens, cyberattacks will explode in both volume and unpredictability.

To bridge the gap, Anthropic recently rolled out Claude Security, an off-the-shelf product powered by Claude Opus 4.8 that scans code and recommends patches, while also sharing its internal vulnerability-finding tools with trusted security teams upon request.

Moving from hunting to patching

Finding 10,000 critical flaws sounds like a win, but it quickly created a massive secondary headache: human maintainers are drowning in reports.

The real chokepoint in cybersecurity right now is verifying, disclosing, and actually patching code at the speed AI uncovers issues. To solve this, Glasswing partners are now using Mythos Preview to write code patches, run pre-release security sweeps, conduct simulated penetration tests, and even rewrite legacy code into memory-safe languages.

Anthropic says it is actively collaborating with third parties to streamline disclosure and patching for open-source maintainers, while preparing to scale up its Cyber Verification Program to let vetted teams use Mythos capabilities for specific defense tasks.

Why it matters

For developers and system administrators running critical systems, this is a clear wake-up call. The era of "security through obscurity" is officially dead; if an unreleased AI can uncover ten thousand critical flaws in a few weeks, malicious actors soon will too.

Strategically, Anthropic's move is clever. By gatekeeping Mythos Preview behind strict vetting while proactively offering it to high-risk infrastructure operators, it is trying to engineer a structural advantage for defenders before competitive parity hits the market.

However, we think this strategy faces a precarious reality check. Defenders must secure every single line of code, while attackers only need to find one overlooked zero-day. If rival models with zero safeguards emerge in the coming year as Anthropic predicts, the race between automated patches and automated exploits is going to get uncomfortably tight.