Anthropic’s Mythos AI is Finding 10,000+ Exploits: Is Open Source Safe? // Jeremy Nwachukwu

A few weeks ago, I talked about a model named Mythos and how it is changing the cybersecurity landscape. More importantly, I talked about how Mythos was finding serious exploits that had gone undiscovered for over 20 years, and how it was chaining exploits too.

I also talked about Project Glasswing, a team that Anthropic assembled to help keep the world’s software safer and more secure. Now, 5 months into 2026, a lot has happened, and I want to talk about:

the update from Project Glasswing
the other hacks that happened outside of Mythos

So, Project Glasswing finally has an update for us, and it’s bad.

For the past few months, Anthropic has been working together with a group of companies. Some notable ones are Google, Apple, the Linux Foundation, JPMorgan, and many more. Together, they formed Project Glasswing, and now they have an update. According to Anthropic, they found over 10,000 high- to critical-severity vulnerabilities across some of the world’s most important software.

The major problem is not how fast they can find bugs, but how quickly they should disclose and patch a large amount of AI-found vulnerabilities.

But this leads to another question: how did we used to handle vulnerabilities?

Normally, vulnerabilities were found and then disclosed after 90 days, or, if there was a patch, 45 days after the patch. This gave users enough time to get the patch into circulation. But now AI has changed the whole thing. Companies are scared to tell people about security vulnerabilities because if they say too much, they endanger their users since bad actors can recreate and exploit people. But if they do not tell people, then users may not upgrade and may keep a bad version of the app installed, meaning people still get exploited.

What I’m trying to say is that Anthropic said they still cannot tell us what exploits they found with Mythos.

But they did share some achievements Mythos has had. Cloudflare said it found 2,000 bugs, and 400 of those bugs were critical security vulnerabilities. Here are some other opinions and reports:

The UK’s AI Security Institute reported that Mythos is the first model to solve both of their cyber ranges, which are simulations of multistep cyberattacks, end to end.
Mozilla found and fixed 271 vulnerabilities in Firefox 150.
XBOW, an independent security platform, reports that Mythos is a significant step up from other models.
Two new academic benchmarks, Exploit Bench and Exploit GYM, show it at the top of the charts.

But despite all these accomplishments, Anthropic claims that companies like Palo Alto, Microsoft, and Oracle are finding and fixing vulnerabilities faster now.

Other hacks

But this is not the only thing that happened this year. Between January and now, I found over 10 major exploits. Let me list them:

January 2026

Target — source code and developer docs stolen.
BreachForums — the hacker forum itself was breached.
ICE — sensitive staff data leaked.
Monroe University — data breach affecting students and alumni.
Under Armour — customer dataset leaked on a hacking forum.
State DHS incidents — two separate DHS data incidents affecting about 1 million people total.

February 2026

Epstein files release — sensitive data leak from badly redacted files.
Conduent — breach updates tied to an earlier intrusion.
Abu Dhabi Finance Week — exposed passports and IDs in an unprotected cloud server.
FICOBA — French national bank account registry breach.

March 2026

Foster City — ransomware-style cyberattack on municipal services.
Russell Cellular / Verizon retailer — retail customer database uploaded to a hacker forum.
GuardDog Telehealth — records accessed under false pretenses and sold.
Intuitive — targeted phishing attack led to data access.
Sears Home Services — 3.7 million records exposed.
University of Hawaiʻi Cancer Center — files exfiltrated.
Crunchyroll — 100GB of user data extracted.

April 2026

Vimeo — breach tied to compromise of a monitoring provider.
ADT — ransomware attack exposed customer data.
Medtronic — ransomware attack exposed millions of records.
Rituals — customer breach disclosed in April.
Autovista — ransomware attack disrupted services in Europe and Australia.
Inditex / Zara — ransomware claim involving millions of records.
7-Eleven — data breach reported in April roundups.
Carnival Corporation — breach reported in April roundups.
KelpDAO — suspicious activity forced a pause in operations.
Seiko — breach reported in April roundups.
Amtrak — breach reported in April roundups.
Kemper Corporation — breach reported in April roundups.
BePrime — admin-account breach exposed credentials and data.
Canada Life Assurance — ransomware attack, with millions of records claimed exposed.
France Titres — government ID-related exposure reported.

May 2026

Canvas / Instructure — major education-platform breach; Reuters reported the company reached a deal with the hacking group.
NVIDIA GeForce NOW — partner breach exposed user and database data.
Harvard — affected by the Canvas breach.
Stanford — affected by the Canvas breach.
UC Berkeley — affected by the Canvas breach.
National University of Singapore — affected by the Canvas breach.
Government of Guam — cPanel zero-day attack disrupted government websites.
San Diego Community College District — cyberattack affected student systems.
ALS Limited — unauthorized access to IT systems.
Canonical / Ubuntu — DDoS attack caused temporary outages.
Trellix — attackers accessed code powering security tools.

What I excluded

Pure outages with no confirmed breach details.

These are just some of the hacks, among many more, that happened this year. And as you can see, that is a lot.

But what does this mean for you?

With models like Mythos, which as of the time I’m writing this has not yet been released but is suspected to release next month in June, it means nobody is really safe. There are supply chain attacks every month. It is best to live your life as if you were already hacked, because you do not know what package has been compromised.

Right now, I recommend that you use the 3-2-1 rule of storage: 3 copies of your data, stored on 2 different types of storage media, with 1 copy offsite. This helps make sure that if you get hacked, you can still recover your files.

Try to keep yourself anonymous online, and do not put your personal data out there. But that is how it affects you — how does it affect open source?

Open source vs AI

So now that we have talked about how AI is being used in cybersecurity, how does it affect open source?

Well, I want to talk about a recent exploit: a 732KB Python file that, if rewritten in Linux, could elevate any user to root privileges. And you may be asking why I’m telling you this. The reason is: do you know the Linux kernel your database is on, or your server, or what version of Linux it is running?

If you do know, and you have updated to a version that is not affected by this exploit, then you are a lucky duck. But the rest of us, who have a life, do not always know what version of Linux our systems are running. So right now, what you are running could be compromised if it is on a version of Linux affected by this exploit.

If you want to know what version has the exploit, I do not know. But it has been a month since the exploit came to light, so by now people should know.

Why do I bring this up? Because open source has great distribution, but not always a great update rate. And update rate is the most important thing to people now, because exploits are easier to find and report. Those exploits have become less desirable to ignore.

So you need to be on the most stable recent version. That does not mean upgrade immediately no matter what — wait a day or two before upgrading anything. It does not matter if it is from npm, pip, apt, yay, or brew; it is all vulnerable to supply chain attacks, so we have to be careful.

And if you see a security vulnerability in software you use, report it and try to help get it fixed. If that means doing a PR, do it. Just give useful suggestions, but do not spam open source projects with random pull requests that are wrong, because then you may actually slow down the maintainer instead of helping. Approach things with caution.

Conclusion

So my conclusion is this: we live in a world where security vulnerabilities are coming out like an infinite receipt-printing machine. Keep your data backed up and keep it safe.

And remember this:if your password is simple, you can get pwned.If it is complex, you can still get pwned.And if you do not know your password, you are still getting pwned.

Honestly, I do not know. I hope that in another couple of weeks I can say this article is irrelevant and all software is secure forever. I wish. But the current landscape is rough.

Reach out on X at JemoLife0213 and tell me your opinion. Am I wrong, or are you also scared?

And remember: stay coding.