Project Glasswing: AI finds the flaws, patching lags
On 7 April Anthropic gave an unreleased AI model to Apple, Microsoft and others to find software flaws. Finding is getting cheaper; fixing sets the pace.
Published · 4 min read

On 7 April Anthropic announced Project Glasswing, which gives an unreleased model, Claude Mythos Preview, to a group of launch partners that includes Amazon Web Services, Apple, Google, Microsoft and the Linux Foundation, and to more than 40 other organisations that build or maintain critical software, so that they can find and fix the flaws in it. Anthropic says the model had already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser, and that it does not plan to make the model generally available. For an office, the news is not the model. It is what happens when finding a flaw gets cheap and fixing one does not.
Its examples were old flaws in heavily tested code. One had sat in OpenBSD, an operating system known for its security, for 27 years. Another, in the FFmpeg video library, was 16 years old, in a line of code that automated tests had run five million times without catching it. Both have since been fixed. Anthropic's own summary is that the cost, effort and expertise needed to find and exploit software flaws have all dropped sharply, and that the same ability is invaluable to the people fixing them.
The bottleneck moves to patching
Anthropic's first progress report, on 22 May, showed where that leads. Anthropic and its roughly 50 partners had used the model to find more than ten thousand high- or critical-severity vulnerabilities. Across more than 1,000 open-source projects the model flagged 23,019 possible flaws, 6,202 of them high or critical by its own estimate. Of about 530 serious bugs disclosed to maintainers by then, 75 had been patched, and some maintainers had asked Anthropic to slow down because they needed time to design fixes. The report put the shift in one line: progress used to be limited by how quickly flaws could be found, and is now limited by how quickly they can be verified, disclosed and patched.
The fixes that do land are arriving in bulk. The same report points to a Palo Alto Networks release with more than five times its usual number of patches, and to Microsoft expecting its count of new patches to keep growing for some time. Its advice to the people who run systems was short: shorten patch testing and deployment, and get the basics right, such as multi-factor authentication, hardened default settings and logs kept for detection.
What it means for an office's own systems
Every system an office runs sits on code somebody else wrote: an operating system, a database, a web framework and many libraries, much of it open source. If the report is right, fixes for all of it will keep coming faster and in larger numbers, and a fix is worth nothing until it is installed. Even a modest set of internal systems, a member portal, an approval workflow and a reporting tool, is a long list of parts, and the list is what makes a flood of fixes manageable, because it tells you quickly which advisories touch you and which do not.
The systems that worry me are the ones nobody maintains any more. A product whose vendor no longer issues updates will not get these fixes, and the flaws in it are getting cheaper to find. That is usually the moment to move the records off it.
Five things to put in place
- Keep an inventory of what each system runs on, with versions: operating system, database, framework and main libraries. Without it, every advisory is a guess.
- Agree a patch window in writing with whoever maintains each system, measured in days for critical fixes rather than in quarterly visits.
- Ask each supplier how they would handle a surge of fixes, and who tests an update before it reaches your live system.
- List the systems that no longer get updates from anyone, and give each one a date to be replaced or moved.
- Do the basics that do not wait on a patch: multi-factor authentication on every account that can change a system, and logs kept long enough to look back through.
If one of your systems is on that fourth list, getting its records off safely is the work behind Move it, which starts with a two-week trial run. The fit check is the place to start.
Sources
- Anthropic, Project Glasswing: Securing critical software for the AI era, 7 April 2026
- Anthropic, Project Glasswing: An initial update, 22 May 2026
- eWeek, Anthropic's Claude Mythos Flags 23K Potential Open-Source Security Flaws, 26 May 2026
Enrique Lacambra III builds internal business systems from Cainta, Rizal.
Tell me what is broken and I will tell you if it is worth building. A 20-minute call, no charge, and I will say so if you do not need me.
