Project Glasswing

Almost every company runs on open-source software, usually far more of it than the inventory suggests. A modern application pulls in hundreds of transitive dependencies, and the projects behind them are frequently maintained by a handful of volunteers. Project Glasswing does not create that exposure, but it makes its size measurable for the first time.

The honest reading is that the asymmetry has narrowed on the discovery side and widened on the response side. A frontier model can now audit a codebase that no human team would have the time to read. That helps defenders enormously, and it means an attacker with comparable access is no longer limited by how many skilled people they can hire.

What follows in practice is unglamorous and well understood: know what software you actually run, keep a usable software bill of materials, and be able to answer within hours rather than weeks whether a newly published vulnerability affects you. Organisations that already have that in place will experience the coming years as manageable. Organisations that do not will experience them as a series of surprises.

There is a genuine opportunity here as well. The same models that find vulnerabilities can review your own code, and they do it at a cost that puts serious security review within reach of mid-sized companies for the first time. Security review used to be a budget line only large enterprises could justify. That is changing.

Project Glasswing is also worth studying as a governance model, independently of its findings. Anthropic did not simply publish a model that can write exploits. It built the capability into a restricted variant, lifted the cybersecurity safeguards only for verified cyber defenders and infrastructure providers, and kept the generally available model unable to do the same work. You can read the details in our topic on Claude Fable 5.1, where the Fable and Mythos split is explained.

The pattern generalises. Inside a company, the question is the same one at a smaller scale: which capabilities should be available to everyone, which ones require a verified need, and who decides. An AI assistant that can read every contract in the company is useful to legal and inappropriate for an intern in marketing. The answer is not to block the capability but to gate it on role and documented purpose, which is exactly what a serious AI platform has to make possible.

For regulated industries this is not optional. The documentation duties in the EU AI Act point in the same direction: know what your systems can do, know who may use them, and be able to show what happened.