Meanwhile on GitHub Claude Code has over 5k bug reports, currently open.
Some good debunking here: https://www.flyingpenguin.com/the-boy-that-cried-mythos-verification-is-collapsing-trust-in-anthropic/
This fluff piece has quite the pie-in-the-sky attitude toward the blue-teaming applications of AI.
Some commentators predict that future AI models will unearth entirely new forms of vulnerabilities that defy our current comprehension, but we don’t think so.
How reassuring.
The defects are finite, and we are entering a world where we can finally find them all.
Could’ve said the same thing when enterprise anti-malware came onto the scene decades ago, but the reality was it was just another vector for the arms race between the red team and the blue team. The author seems to put a lot of stock in the whole “the blue team has access to these AI tools that the red team doesn’t currently have access to” argument, which kinda ignores the fact that that reality is simply not going to last.
I could be wrong, but any article suggesting “zero-days are numbered” doesn’t pass the smell test.
The author seems to put a lot of stock in the whole “the blue team has access to these AI tools that the red team doesn’t currently have access to” argument
I didn’t read it like that. I think the point was that the red team had an edge over the blue team (by being able to spend a lot of effort on a single exploit), so when both teams have access to these same tools, it’ll be more of an equal fight.
I could be wrong, but any article suggesting “zero-days are numbered” doesn’t pass the smell test.
Yeah, you’re right.
The real story is that it is a bit better at finding bugs. Calling them zero-days and implying there’s some major security implications is just to build hype.
It was able to chain a few of the bugs together to create a RCE exploit in a weakened browser, it’s interesting but don’t go to your fallout shelter just yet.
We’ve led the industry in building and adopting Rust
Yeah, then you fired the team to pay the CEO a few million more.
Defenders finally have a chance to win, decisively
I’m curious how it will turn out to be in a long term. Are we going to have safer software? Because not only defenders will have a powerful tool, but attackers too. But at the same time, number of bugs is finite… Can we in theory one day achieve literally zero bugs in codebase?
It does seem advantageous to the defender.
Another factor Mozilla didn’t mention (and that Anthropic wouldn’t like to emphasize) is that major LLMs are pretty similar. And their development is way more conservative than you’d think. They use similar architectures and formats, train from the same data, distill each other, further pollute the internet with the same output and so on. So if (for example) Mozilla red teams with Mythos, I’d posit it’s likely that attacker LLMs would find the same already-patched bugs, instead of something new.
…So yeah. I’d wager Mozilla’s sentiment is correct.
Add to that that AI is pretty good at copying from pre-existing knowledge (like a database of known vulnerabilities) and not good at generating novel ideas (like discovering a new vulnerability), and the scales are further tilted in the defenders’ favor.
Eh, I don’t totally agree. AI can discover novel exploits that aren’t already in some database, and likely have in this case.
I’m just saying the operating patterns between different LLMs are more similar than you’d expect, like similar tools from the same factory.
Are we going to have safer software? Because not only defenders will have a powerful tool, but attackers too.
Probably not safer software, but the window of time for a bug being known and exploitable will be shortened greatly. Instead of 0-days, we might have 0-minutes.
That’s assuming these ridiculous AI systems are rolling deployments that fast, so maybe that idea’s nonsense.
You can achieve zero bugs through liberal use of rm.
You can achieve the same effect with a hammer
deleted by creator
Cyber security in general is going to get interesting. Breaking into protected systems often requires more patience than expertise. Attackers often get detected when they take short cuts because of laziness and overconfidence. AI agents have unfathomable patience and attention to detail.l
I don’t really agree with the attention to detail part from my experience. AI agents love to take shortcuts from what I’ve seen, and you have to pay a lot of attention to what they’re doing to make sure they do the right thing.
AI will be good at scaning for known vulnerabilities, but patience and attention to detail? Not in my experience. I use agentic coding agents for work and they are getting better, but they still will regularly get stuck in a loop of running into a bug when running tests, attempting to fix the bug in a stupid way, still erroring, trying another stupid fix, trying the first stupid fix, and so on until a human intervenes. They may be patient (as long as you pay for more tokens), but they aren’t using their time wisely.
AI tends to use the “throw shit at the wall and see what sticks” approach. It’s getting better at writing maintainable code, but it still will generate more-or-less spaghetti code with random unused or deprecated variables, crazy unnecessary functions, poor organization, etc… and requires lots of testing before producing something functional. Which is fine in an environment where you can iterate and clean things up. But as an attack vector, if you need 58 attempts to fully realize a vulnerability, in most secure environments you’re going to get detected and blocked before you finish.
I don’t disagree on the current state. However, it’s not hard to foresee that attack tools will be developed that can maintain “attention” on an attack for days or weeks at a time with privately run agents. I’m sure they are out there already to some degree.
They have attention to detail, just not the right details. It’s super easy for them to get lost in a never ending train of tangents.
It is theoretically possible by using formal verification. Which is getting easier due to lean. But still impractical.
Not zero bugs, but it should help. A benefit for defenders is that they can use AI review on code before they make it public or release it in a stable release
How many vulnerabilities would’ve been found if we had spent several million dollars on human security researchers though?
Slopzilla Slopfox 🙄
This isn’t going to end well.
If it’s finding valid vulnerabilities then it’s just another tool like static analysis, fuzzers and sanitizers. There definitely seems to be a difference in quality compared to earlier generations that were behind the sloppy avalanch of reports.
Instead of 271 vulns, it was more like 2 vulns, found 271 different times.
Here’s a o good article about it https://www.flyingpenguin.com/the-boy-that-cried-mythos-verification-is-collapsing-trust-in-anthropic/
There’s a difference between using AI to apply fixes for problems, and using AI to find problems that you didn’t know about.
Mythos does the latter, not the former.
It’s only a matter of time until they decide it should do both and it makes a mess. Calling it now.
Yes, the sky is falling, AI is ending the world, slopslopslop, etc
We know the bit.
That doesn’t make sense. Don’t the attackers have the same tools?
Mythos Preview is better at finding real vulnerabilities than existing public models and, for now, only a few have access to it.
I’m aware (unfortunately) of the marketing claims and even if they might be true, as you say it is “for now”. So if it’s only temporary for that arm race, especially if held by a company who leaked its own code just days ago, then I have a hard time understanding why ‘zero-days are numbered’ because this title claims the dynamic itself is gone. That’s now my understanding, especially if other models are just marginally (which is hard to prove with models, finding proper metrics) worst than it.
See comment that shared https://techcrunch.com/2026/04/21/unauthorized-group-has-gained-access-to-anthropics-exclusive-cyber-tool-mythos-report-claims just few hours ago, and that’s not even sophisticated.
Anthropic and OpenAI have multiple times used this arm race rhetoric before and it worked. Their models are supposedly “too dangerous” to be released thus consequently they have to control access.
It might be true but so far what we have witnessed is that roughly equivalent models get released by others merely weeks or maybe months after, sometimes open, but the “moat” never lasted long so I’m questioning why it would be different this time.
Actually untrue. The only thing mythos added was an automatic way to exploit vulns that other models also find. I read a good article on mastodon about it. I posted it elsewhere in the thread but also here https://www.flyingpenguin.com/the-boy-that-cried-mythos-verification-is-collapsing-trust-in-anthropic/
for now
Not right now, thats the whole thing
