Field NotesStory5 min read
Anthropic's Mythos model finds decades-old bugs, and small companies need to patch faster
Anthropic's new model found thousands of serious software flaws and won't be released. Here is what that means for a company of 20 to 200 people.
Adam Blazer
Fractional Chief Intelligence Officer
Last Tuesday, Anthropic announced a model it says it won’t sell to the public. Claude Mythos Preview found thousands of high-severity security flaws in major operating systems, web browsers and other widely used software, according to the company. Instead of a general launch, Anthropic gave access to a small group of large technology and security companies through a program called Project Glasswing.
Newton Cheng, who leads cyber work on Anthropic’s Frontier Red Team, told VentureBeat: “We do not plan to make Claude Mythos Preview generally available due to its cybersecurity capabilities.”
If you run a 50-person company, you won’t be getting access to this model. It still changes your year.
What Mythos found
The examples Anthropic shared are the kind that make security people sit up. According to VentureBeat, the model found a 27-year-old bug in OpenBSD that let an attacker remotely crash any machine running it. It found a 16-year-old bug in FFmpeg, a widely used video library, in code that automated testing had run through five million times without catching the problem. It also chained together several Linux kernel flaws to go from ordinary user access to full control of a machine. All three have been reported and patched, and Anthropic says other findings are still being fixed.
that an OpenBSD bug went unnoticed before Mythos Preview found it
On CyberGym, a benchmark for this kind of security work, Mythos Preview scored 83.1 percent, against 66.6 percent for Claude Opus 4.6, a model any company can use today.
- Claude Mythos Preview83.1%
- Claude Opus 4.666.6%
The launch partners include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. Anthropic also opened access to more than 40 other organizations that build or maintain critical software. It committed up to $100 million in usage credits and $4 million in donations to open-source security groups. After the credits run out, participants pay $25 per million input tokens and $125 per million output tokens.
Why a model you can’t use still affects you
Three things follow from this announcement, even for a company that will never touch Mythos.
A wave of patches is coming. If one model can find thousands of serious flaws, the companies behind your operating systems, browsers and core software will be shipping fixes for them over the coming months. Each fix is only useful once it is installed. A company that updates laptops and servers whenever someone gets around to it will be running known-vulnerable software for longer than it should.
Old software is now easier to break. The bugs Anthropic highlighted were 16 and 27 years old. They sat in code that people assumed was safe because it had been around forever. VentureBeat reported Cheng’s warning that these capabilities could spread to people who wouldn’t use them carefully. Whatever the timeline, the old assumption that an obscure, aging system is too boring to attack is getting weaker. The old WordPress site, the forgotten file server and the accounting package nobody has updated since 2019 are the systems to worry about.
The attackers are already busy. The same day as the Glasswing announcement, The Register covered the FBI’s annual internet crime report. Reported losses hit a record $20.87 billion in 2025, and complaints topped one million for the first time. The FBI counted 22,364 complaints involving AI, with more than $893 million in losses, and said that figure is likely an undercount. Most of those losses come from fraud aimed at people. Faster discovery of software flaws only adds to the pressure.
What changes for a 20 to 200 person company
Most companies this size don’t have a security team. They have an IT person, an outside provider or a founder who handles it. That is fine. The response to this news is mostly basic hygiene done on a schedule.
Know what you run. Make a list of every system the company depends on: laptops, phones, servers, the website, the store, the accounting software, the plugins on all of them. You can’t patch what you don’t know you have. Include the things people set up on their own, like a shared drive a team created on a personal account or a plugin someone added to the website years ago.
Turn on automatic updates wherever you can. Operating systems, browsers and most business software can update themselves. For anything that can’t, set a weekly time for someone to check and install.
Retire what nobody maintains. If a piece of software no longer gets security updates from its maker, plan to replace it this quarter. An old system that works perfectly is still a liability if nobody will ever fix its flaws.
Ask your vendors one question. For each critical software provider, ask how quickly they ship security fixes and how they will tell you. The answer tells you a lot about whether they are ready for a faster cycle.
- Every system is on one list
- Automatic updates are on
- Old software has a retire date
- Vendors named their patch speed
- One person owns weekly updates
Look at the code you own. If an agency or a contractor built your website or a custom app years ago, it hasn’t been reviewed since. The models available to everyone today aren’t Mythos, but they can already review code. Have whoever maintains it run a security review this quarter.
There is also a broader point about how frontier AI is arriving. Mythos went first to the largest companies, under controlled terms. That may become more common for the most capable models. For a smaller company, that means the gains reach you through your vendors and the open-source software you rely on, not through your own account. Choosing vendors who are part of that work, and who patch quickly, matters more than it did a month ago.
It is also a reminder of how quickly the tools are moving. A model that finds flaws testing missed for 16 years is a strong signal of what AI coding and review tools will do over the next few years, for defenders and for everyone else. If your company writes any software of its own, even a small internal tool, now is a good time to get into the habit of having an AI review every change for security problems before it ships, with a person deciding what to fix.
What to do this week
- Write a one-page list of every system and piece of software the business depends on, with who maintains each one.
- Check that automatic updates are on for every company laptop, phone and browser.
- Mark anything on the list that no longer receives security updates, and set a date to replace it.
- Email your three most important software vendors and ask how fast they ship security fixes and how they notify customers.
- Put a weekly 30-minute slot on someone’s calendar to install updates that don’t apply themselves.
None of this needs a new tool or a new hire. It needs one person with a list and a recurring calendar slot.
Sources
- Anthropic says its most powerful AI cyber model is too dangerous to release (VentureBeat, April 7, 2026)
- Project Glasswing (Anthropic, April 7, 2026)
- Cybercrime losses reach all-time high (The Register, April 7, 2026)
