Mythos And Mayhem: Knowing The Fix Is Not Enough

The software vulnerability problem is much worse than you might think. When I was trying to think of how to explain it, a silly metaphor popped into my mind based on a recent experience. I can never resist a silly metaphor...

The Plumbing Metaphor

After a month of below-freezing temperatures this spring, I noticed something strange: running the cold tap in the bathroom long enough brought in water from outside the house — which was much colder than usual — leaving a slow drip when I turned it off. The drip would stop after a few minutes once the pipes warmed back up, so I had time to try to fix it myself.

Actually, I didn't discover the drip - my cat did. Drips, it turns out, are super fun! You can try to bite them as they fall. You can also stick your face upside-down under the faucet so that drips go up your nose. This produces a very impressive sneeze! Like I said, super fun.

I did what any reasonable person does these days: I took a picture of the faucet and described the problem to AI (Perplexity, in my case). Within minutes I had a diagnosis — the likely culprit was a cartridge, an internal valve assembly inside the cold water handle. This was the first I'd ever heard of a cartridge in a sink handle. Sheltered life, I guess. The good news was, chances were good that if I replaced the cartridge it would stop the drip.

Armed with that information, I disassembled the faucet, took pictures of the cartridge, and went hunting for a replacement. I spent hours wrestling with the genuinely terrible search interfaces of every hardware and home-improvement store I could find. I eventually located what looked like the right part. I ordered it and waited a few days for it to arrive. It didn't quite fit. Close, but not close enough — and after a bunch more research, I discovered that our faucet was more than twenty years old, and that neither the manufacturer of the faucet nor anybody else makes a compatible cartridge anymore. All modern cartridges were too new. The only path forward is replacing the whole fixture.

By the time I figured that out, the weather had warmed up and the drip had stopped, and I had decided to call in an expert. Our plumber (who is excellent) will arrive, look at the faucet, look at me, and suggest that this is something a functional adult should have been able to handle on his own to save money. He's right, but of course he doesn't know about "the Portland Incident". I know my limits.

The moral, before we get to the part where the stakes are considerably higher than a leaky faucet:

• I found the problem (actually my cat found it first)
• I figured out the likely fix (replace the cartridge), with help from AI
• Version incompatibility meant the likely fix didn't exist, and fixing it right meant replacing a much larger and more expensive thing
• And that is nothing compared to fixing software

Mythos, and Why Anthropic Didn't Sell It

In April 2026, Anthropic announced that they had built a new AI model called Mythos — and that they were not going to release it for general use.

Mythos is their most capable model to date, and its area of unusual strength is cybersecurity. The model can identify and exploit security vulnerabilities across major operating systems and web browsers. Anthropic says it has discovered thousands of serious flaws, including some that had gone undetected for decades. In testing, Mythos was able to autonomously exploit Firefox vulnerabilities with nearly 75% success. Previous models managed close to zero.

Instead of selling it, Anthropic gave access to a few dozen leading tech and cybersecurity organizations — along with $100 million in free usage credits — specifically to find and fix vulnerabilities before a tool this powerful ended up in the wrong hands. They already had evidence that hostile actors had tried to weaponize earlier models in attacks on financial institutions and government agencies.

The predictable response in some corners of the internet was that this was a marketing stunt. Look at us being so responsible. Steven Adler, writing at Clear-Eyed AI, makes the obvious-once-you-say-it counterargument: companies do not typically market products by emphasizing the catastrophic harm those products might cause. They also don't generate revenue by building expensive things and then deliberately not selling them. And they don't subsidize customer usage if the goal is profit.

Adler also cites survey data suggesting that top AI researchers put the median probability of AI causing human extinction or civilizational collapse at somewhere between 5% and 10%. Anthropic's founders left OpenAI and pledged 80% of their shares to charitable causes. Whatever you think of their judgment, the evidence doesn't point to people who are mostly doing this for the marketing.

Knowing the Fix Is Not the Same as Applying It

Here is where the faucet analogy stops being charming and starts being useful.

Even if Mythos finds thousands of security vulnerabilities — even if every one of those vulnerabilities has a known fix — that does not mean those fixes can be applied. Because modern software doesn't run in isolation. It runs in ecosystems of dependencies, and those dependencies have dependencies of their own.

When you install a software package, you're not just installing that package. You're installing everything it depends on, and everything those things depend on, in a chain that can extend surprisingly far. A 2024 study on transitive dependencies found that in the Java/Maven ecosystem (for example), declaring a single dependency pulls in an average of about 25 additional packages. A Maven project that declares five direct dependencies ends up with an average of 80 total packages installed. And because of how this compounds, the correlation between the number of direct dependencies you declare and the total packages you end up with is weak enough to be nearly useless — you genuinely can't predict your exposure from looking at your own imports.

The security implication is direct: a single vulnerable package can affect more than 24 times as many projects as direct dependency counts would suggest.

This is what's known as dependency hell, and it creates a specific, nasty version of my little plumbing problem:

Your application (A)
  └── depends on Library B (v2.3)
        └── depends on Component C (v1.4)
                         ↑
              Security fix exists in C v2.0
              But C v2.0 breaks Library Bs API
              ** And nobody has maintained Library B in three years **

You can't upgrade C without breaking B. You can't fix B because no one's working on it. So A is stuck depending on a version of C with a known vulnerability, and there is no clean path out — only expensive ones. You'd need to either replace B entirely, fork and update B yourself, or rewrite the parts of A that use B. None of that is fast. Any often, there are way more than 3 levels of dependencies.

Now multiply this by the scale of what Mythos found. Thousands of vulnerabilities, scattered across codebases that are themselves tangled in dependency chains that nobody fully understands. Each fix is potentially a thread that, when pulled, unravels something else.

The Ground Is Already Shifting

The window for getting ahead of this is not large. Anthropic has Mythos. Other labs are close behind. The knowledge of how to find these vulnerabilities at scale already exists; the only question is how soon a version of it ends up being released by someone who isn't handing out $100 million in free credits to fix things first.

Finding the bugs is the tractable part. We are surprisingly good at that now. What we are much less good at is the compatibility problem at scale — the version incompatibilities, the unmaintained packages, the dependency chains that turn a straightforward fix into a multi-month engineering project, multiplied across thousands of vulnerabilities, in systems that are still running in production, in organizations that don't have the time or budget to do it right.

It's a very big problem, and not one we can see clearly or predict.

The city bus in the photograph didn't fall into that Pittsburgh sinkhole because nobody knew sinkholes could happen. It fell because nobody knew that particular sinkhole was there, underneath something heavy that couldn't stop in time.

We, as an industry, have a brief moment in time to fix the most critical issues before things really fall apart.

Thanks to Anthropic, we have a warning. And that's nothing to sneeze at.

Photo: daveynin, CC BY 2.0, via Wikimedia Commons