"My personal conclusion can however not end up with anything else than that the big hype around this model so far was primarily marketing. I see no evidence that this setup finds issues to any particular higher or more advanced degree than the other tools have done before Mythos. Maybe this model is a little bit better, but even if it is, it is not better to a degree that seems to make a significant dent in code analyzing."
It's a good reminder for us all that the competition in this space is rough and lots of more or less subtle marketing is involved.
This is roughly what I was assuming but of course the big caveat here is that they were already using the existing LLM driven tooling on an extensively audited codebase.
So while anthropic's marketing may be hype there just wasn't much left to find, a point he makes in the blog post.
Whether it's a big step forward for other kinds of projects is difficult to tell, but this highlights that everybody should be using AI code review tools to audit their existing code today, and not everybody is.
Anthropic using marketing to convince people their models are more advanced, better built, or that AI is a threat that needs to be regulated because only they have the answer? I’m shocked.
More seriously, so far I haven’t seen much indication that Mythos is more than Opus with a security focused code analysis harness. That said, the fact it can find these bugs in an automated fashion is the more important takeaway outside of the hype.
I’m curious what the error rate is on the detections, because none of that means much if it is wrong 90% of the time and we are only hearing about the examples that are useful marketing.
>> Anthropic using marketing to convince people their models are more advanced, better built, or that AI is a threat that needs to be regulated because only they have the answer? I’m shocked.
I remember when OpenAI was saying GPT-2 was too dangerous to release.
I remember when there was a guy at Google years a few years ago that was convinced that they had an internal, sentient creature in their labs (I think maybe 4 years ago?)
If I’m not mistaken, after the media cycle, he lost his job for breaking confidentiality.
That was the opposite of marketing, Google really didn’t get how to turn this into a product until ChatGPT happened.
I always thought he was fired for making crackpot statements to the press in reference to his professional capacity, and thus creating bad PR and embarrassing spectacle for his employer. Seems like legitimate reasons to me.
An interesting question now is whether he had standard mental health issues, or if he was an early example of AI psychosis or whatever we call people who are falling in love with their AI chatbots because they tell them how smart they are.
Considering Richard Dawkins has recently succumbed to the same delusion it is a reminder that no matter how intelligent someone may otherwise be, we are all human and have certain tendencies and blind spots; anthropomorphizing non-entities being one of those.
Richard Dawkins is 85 to be fair, just like Bernie Sanders is 84 when he made similar comments. The other guy worked on Google's AI safety team where one would assume he'd have a basic grasp of how the technology works.
Given how much money is on the line, it would be gross negligence if anything came publicly out of the CEO's mouth or is otherwise published by the company that's not marketing.
Evidences: 10 years ago, when I interviewed Baidu AI with Andrew Ng and Dario, Dario is the kind of person is pure-hearted to the point being ideological. Given Dario's successful career so far, that essence has gradually grown into a conviction, and surrounded by a purposely built team which amplifies his ideology.
Humans are very convenient creature, a rare few small fraction of them are no doubt the master of convenience: they morph their mental manifold without a hint of contradiction in their own mental mechanisms.
Mythos put Anthropic back into the White House’s good graces. It also branded Anthropic as badass, something their softener image probably needed to win government contracts.
Maybe it wasn’t marketing. But the product’s configuration, and how Anthropic talked about and released it, sure as hell played beautifully. (The timing, while Musk and Altman are distracted with each other, also couldn’t have been better.)
These things are layered. They are great scientists, smart people, etc.
Things change when you’re running a business like Anthropic, especially as the CEO. You have a responsibility to shareholders, and you just need to play the game.
Anthropic chose a great angle: focus on professionals / enterprise, safety, etc. Those can both be done by a genuine desire to make great technology, and for business purposes require you to position yourself in a bit “better” way than reality.
Just look at what their strategy is with Mythos, it’s almost perfection: the “it’s not ready to be released to the public” angle hits all the marks: they care about responsibility / safety, they have “the best” model, and “LLMs are dangerous, but we, as the guardians, can be trusted”. This also helps the industry as a whole with regulation: if they’re being constrained, China will develop even more dangerous models.
This is a result of how smart people treat business, it’s PR perfection, especially given how much the whole industry is talking about it.
(Yes, they fail in other PR areas, but that’s a different discussion)
That's an odd definition of "intentional". Evolution has filtered for people with certain views and the marketing has just emerged from their actions. ... So?
A deadly virus (naturally occurring one let's say) wasn't created intentionally. Evolution selected for it. It's still bad and kills people. Doesn't make it nice because of lack of intention.
I'm not sure if that distinction is important, since what you've described less charitably synonymous with the phrase "Dario is delusional, and has surrounded himself with yes-men, so outlandish marketing gets published as a side effect".
Whether the person doing the marketing was sincere about it or not is immaterial, since marketing is experienced almost entirely by the people consuming it, and not the people communicating it. What matters is if the audience is sincerely concerned by the message, and it's transparently the case that they were sincerely concerned by it.
Curl simply isn't a good data point. It's one of the most picked-over codebases in existence with extensive security testing practices. All the researchers using not-quite-Mythos models have had plenty of time to report bugs up to this point. Daniel may be right that Mythos hasn't been a game changer for curl but the preconditions are different for virtually any other codebase. Perhaps the real marketing here is his own modesty about curl's maturity.
Curl uses all sorts of tools, including AI tools to find bugs. These tools, according to the article found hundreds of bugs including a dozen CVE.
Mythos found one vulnerability. It means the Mythos is just another tool, not the revolution it claims to be.
It is common that when a new tool is introduced that a bunch of bugs are found, with diminishing returns. Mythos finding one vulnerability is consistent to what I would expect for a major update to an existing tool, which Mythos is over existing LLM-based solutions.
The question is how many security vulnerabilities are actually left in the code after all the recent AI attention. Either Mythos is a nothingburger, or it's substantially more powerful but there's nothing left to do. Even a large amount of C can be correct eventually. Curl has the _potential_ to become a good data point maybe 6-12 months from now - if researchers and new tools find many more vulnerabilities then Mythos is proved to be hype. If they don't, then maybe Mythos is overkill for today's curl and its capabilities are better deployed elsewhere (like Firefox, apparently).
I have a hard time believing that Mythos found the only remaining Curl vulnerability. It is possible, but highly improbable.
And it is not overkill, the proof is that it found that vulnerability. It is like saying the new version of some static analyzer with some new rules is "overkill" because it only found only one more bug than the previous version. Deciding whether it is overkill or not is more about context. Using a very expensive model like Mythos for some little used non-critical software is overkill, but for Curl, it absolutely isn't.
If Mythos found loads of vulnerabilities in Firefox but not in Curl, I wouldn't say that's because of Mythos is so good, but rather that with the release of Mythos, they did some testing that could have been done before using the same tools Curl have used.
> Once the end-to-end pipeline is in place, it’s trivial to swap in different models when they become available. Building this pipeline early helped us find a number of serious bugs using publicly-available models, and it also helped us hit the ground running when we had the opportunity to evaluate Claude Mythos Preview. In our experience, model upgrades increase the effectiveness of the entire pipeline: the system gets simultaneously better at finding potential bugs, creating proof-of-concept test cases to demonstrate them, and articulating their pathology and impact.
We see this exact hypetrain every time a new model is released. Mythos simply hasn't lived up to the "we're all gunna die from the flood of vulnerabilities" hype even slightly. Its slightly better than previous models by all accounts, cool stuff
I've seen literally near word-for-word this exact chain of events multiple times previously
I've been running my own security scanning software (disclaimer: now starting a company @ zeroquarry.com) for this, and from what I've seen there's a huge value in prompts + adversarial LLM review. Without adversarial review, you get garbage (as this blog points out: 4/5 basically are nonsense) and with a good prompt, you can use almost any "near frontier" model from my experience as long as the prompt helps with the guardrails or the model doesn't protect in such a strict way
Mythos marketing really leans into that "too powerful to be legal" vibe, much like how PS2s were allegedly banned from North Korea because their chips were basically missile-grade.
Eh... I think he puts the LLM down for his own ego's sake (as would I!). Curl may, next to the Linux kernel, be one of the most heavily audited codebases in existence. The LLM found something he and thousands of others missed. It's not unimpressive.
I'd go out and say the marketing is not subtle. The hype and fanboys/girls are so in line with the marketing that any level of skepticism is seen a an act of defection, but if you look at the words, hyperbole and volume that is used, there is nothing subtle about it.
It's almost Trump-esque - "this model will change everything forever; we are doomed; we are saved; we will all be fired; we will all be rich", etc
That's a pretty good encapsulation of the parallels between the political and the technological: One necessarily thrives upon the other and are inextricable. This moment is a culmination of all the disenfranchisement the bodypolitik have suffered, looking for any possible means of escape or elevation. AI and Trumpism, for their own respective cohorts, are salvation, on offer by different frontmen but ultimately in service of the same system.
They need the hype to pay off way more than we do. So many of us who still write code directly stand to lose nothing of our capabilities if the marketing claims cannot hold water.
I seem to be totally outside the hype bubble, but I have to suspect there is a lot of imagineering and wild extrapolations in the elss technical hype bubbles. I am curious but no enough to go looking.
I'm surprised you say that because it is all over Hacker News. Every single post is co-opted into promoting AI. Try finding a submission with fifty points or more than doesn't have AI or LLM's mentioned somewhere in the comments.
> An amazingly successful marketing stunt for sure.
This. Well done by Antropic.
It even reached the CISO of my small semi-government org in the Netherlands, who slightly panicked at the announced 'tsunami' of vulnerabilities that was coming with Mythos.
Got us some more money and priority with the board, though.
I don't agree with the "no tsunami in sight": if you don't look at 100+ bugs in Firefox and many more OSS projects, bunch of old unseen-before OpenBSD/Linux RCEs, and a few LPE in just 2 or 3 weeks for Linux itself...
IMO, this does not sound like marketing scare, there is spike of vulnerability disclosures - high quality, low false positives - that can be sensed... It feels like we're speedrunning through few-years worth of high quality bug reports in just a few weeks.
Anthropic has is quickly destroying customer goodwill by repeatedly pulling the same stunt. Horrible marketing, imho.
It's an entirely different thing to have the company conduct research on LLMs in general being a cybersecurity threat, instead of going "our new model is just too powerful" and shift the discussion to revolve around that. It's slimey.
Sure, but isn't it a verdict on Mythos compared to other models?
If so, it would still follow. "Most software" isn't analyzed as much as curl, by either other tooling or other models, that might well find close to the same as Mythos did. As such, Mythos then isn't especially/particularly dangerous.
I don't think I understand what you mean, the "not particularly dangerous" comment was in relation to the vulnerability that was found right ? Surely they would know what constitutes a lower severity level.
My guess is that it is in category of "you are holding it wrong". Still worth fixing, but requires very specific user input for example. Or very weird scenario. Or in some less used protocol or flag combination.
Curl is currently receiving a record number of high-quality bug/vuln reports (a rather sharp change from the earlier slop inundation), so it’s not like there’s nothing to find. Many or most of these are presumably found by human experts assisted by AI tools, but if Mythos were truly revolutionary, it should be able to find such issues on its own.
> These tools and the analyses they have done have triggered somewhere between two and three hundred bugfixes merged in curl through-out the recent 8-10 months or so.
If you've just gone through a lengthy analysis of your code with other AI tools, surely it's reasonable not to expect to see hundreds more from a new tool?
It should be possible, unless more bugs are introduced, to eventually get to a state where there are no more bugs in your code.
Process aside, it sounds like Daniel expected to find dozens/hundreds more bugs.
> The single confirmed vulnerability is going to end up a severity low CVE planned to get published in sync with our pending next curl release 8.21.0 in late June
My mind still cannot understand the quality and refinement that's gone into cURL. It really is the perfect example of something done so right, that people barely think twice about.
Easy, it shows what is achievable if there is a high bar for quality in every single line of code that gets commited, reviewed and merged, regardless of the programming language.
However in the days of race to bottom, offshoring for penies, and now LLM powered code generation, this is a quality most companies won't care unless there is liability in place.
Curl and SQLite are my favourite examples of properly engineered, rigourously tested _anything_. It's really philosophical - those projects' contribution requirements demand such rigor, and the maintainers stand by that demand. A non-load-bearing document (not project code) is what makes that possible - very reminiscent of Einstein's thought experiments leading to tangible projects such as GPS or Descartes's belief that all problems can be solved through rational thinking.
Some people must be working on training some models exclusively on high quality OSS code base like curl and SQLite without the noise of low quality training data.
I would do that with 100% local models from scratch.
There is always marketing involved and people should be able to put marketing into perspective.
Also curl in this regard is a open source project, relativly small but critical, well known and used everywhere. Besides image libraries, tools like curl or sudo, su, passwd, etc. would also be my first try.
Mythos is still not known at all what it can do. What does it mean from cost and benchmark pov to have a 10 Trillion parameter model?
Nonetheless, the fact that LLMs got significant better in finding this, better than humans, started to happen half a year ago? so at one point we need to address the elefant in the room and state that today you need to do security scanning additional with LLMs. You need to take this serious.
In worst case, use Anthropics marketing to state that its a must now and something changed.
> What does it mean from cost and benchmark pov to have a 10 Trillion parameter model?
To me it means that we've hit the top end of the S-curve with regards to effects of scaling - if the tool isn't remarkably better despite the scale, then we're firmly in diminishing returns territory.
> Nonetheless, the fact that LLMs got significant better in finding this, better than humans, started to happen half a year ago?
*rolls eyes* regular static analyzers also have been "better than humans" for decades, being better than a human at a specific mechanical task really doesn't mean much. The interesting new thing is the type of potential "fuzzy bugs" described in the article that LLMs are able to identify (a comment not matching the code it describes, uncommon usage of a 3rd party library, mismatch of code and a protocol it implements, or often just generally weird looking code somebody should have a closer look at... this closes a gap in the traditional debugging toolboxes, but shouldn't replace them)
Putting on my tinfoil-hat: Sooo, the guy who runs the test and delivers the report could just have removed the more interesting bugs and delivered those to any three letter agency?
curl's source is public so what would be the gain in the rigmarole? Now if the prompt was "create a patch that inserts a zero-day while fixing a bug" that would be impressive.
I don't know about Mythos but in recent weeks I've noticed Opus is constantly failing to fix things in tsz[0] vs GPT 5.5 can easily churn out fixes that are solid and pass tests. I've stopped paying for Claude for now and all my money is going to OpenAI at the moment. Either Opus is massively nerfed or GPT 5.5 is really head and shoulder higher in terms of very difficult tasks. The last percent of conformance tests in tsz are really really difficult and I've seen Opus bailing again and again. So annoying to waste time and tokens to finally get "this is too involved" or "this requires a multi-week sprint to fix".
From a user’s perspective 4.7 is a downgrade compared to 4.6 . It’s intended to give Anthropic more control about their compute resources and profitability:
I am curious, what kind of work do you use Claude for that sometimes requires hours of working. In my case, I have never seen it go off for more than 10 mins and even that is very rare.
debugging code. I had some issue so I create a plan to root cause that would run the code, change some functions or variables and run again until we get a confirmed answer.
I just work up to that very workflow this morning. I ran last night and finished at around 3am with ~200k tokens spent. Fixed the issue and created a follow up doc for things that it could not verify.
I'm looking forward to trying Mythos run against my 5000-line, instant-finality, quantum-resistant blockchain project and decentralized exchange (an additional 5000 lines). I already ran all the models up to Opus 4.6 and they couldn't find anything.
I routinely used to compile C programs on other compilers to find defects that one or another didn't find. Compiling on Windows vs Linux. You could summarize / minimize it down to compiling it with warning as errors etc but you'd be missing the point.
The point wasn't actual cross-platform portability even though that was a nice side effect. It was to flush out all the weird edge cases.
Edges like security flaws. Buffer overflows are usually platform specific. There are plenty of other ways to find these issues but simply recompiling for a different platform surfaces all sorts of issues.
Voice input works really well for people speaking English with a Swedish accent. I think the accent of most educated Swedes is mostly a case of prosody. For sure there are some sounds we say slightly differently than native English speakers. We often have some trouble with /s/ and /z/, but I don't know, "war and peace", I think that's easily understood.
Source: voice typing this with Swedish vocal chords, and only had to correct "different lives" to "differently", and add /[^\w\s]/.
Android voice input works with kids using both English and native words, here in India. The country runs schools in 25+ primary languages, each with dialects, so a TV/phone with voice input is more marvelous than the nitpicks discussed here.
War and Peace is about 590,000 words. Tiny compared to the full Harry Potter collection (about 1 million words over the 7 books), but long for a single book.
They're referring to the typo in the title, "Piece" vs "Peace".
I also thought they were contending the word count before noticing. Even remarked how I find this a weird metric, given that code is not prose [0], but then I deleted that once I picked up on what's going on.
[0] comparing the output of `wc -w` with the word counts of books I'm reasonably sure will be super off
edit: ran a calc, substituting out symbols (but not underscores), digits, and comments yields a 390K word count compared to the 660K cited. not excluding the comments yields 600K, so more than a third of all words in the sources are comments.
It's a shame he seems to reject the idea of actually diving in and using these tools interactively:
> It’s not that I would have a lot of time to explore lots of different prompts and doing deep dive adventures anyway.
His expertise I think would elevate the results quite a bit. Although if he never uses LLMs, which it reads like he doesn't, I guess it might backfire just as well. Prompting style (still?) does matter after all, certainly in my experience anyways.
He posts about his use of language models a lot on Mastodon[0]. He does lots with language models, but doesn't buy all the way into the hype. I'd say he's one of. most reasonable & balanced voices on the subject of AI use in software today. Happy to use the technology, more than willing to push back on marketing bs.
Won my bet "voted 10 [vulnerabilities] but in retrospect as you are familiar with Claude and such tooling if you already used any of recent model to done some kind of security review then I'd drop to 1 or even 0." https://mastodon.pirateparty.be/@utopiah/116537456780283420
Quote:
"My personal conclusion can however not end up with anything else than that the big hype around this model so far was primarily marketing. I see no evidence that this setup finds issues to any particular higher or more advanced degree than the other tools have done before Mythos. Maybe this model is a little bit better, but even if it is, it is not better to a degree that seems to make a significant dent in code analyzing."
It's a good reminder for us all that the competition in this space is rough and lots of more or less subtle marketing is involved.
This is roughly what I was assuming but of course the big caveat here is that they were already using the existing LLM driven tooling on an extensively audited codebase.
So while anthropic's marketing may be hype there just wasn't much left to find, a point he makes in the blog post.
Whether it's a big step forward for other kinds of projects is difficult to tell, but this highlights that everybody should be using AI code review tools to audit their existing code today, and not everybody is.
Anthropic using marketing to convince people their models are more advanced, better built, or that AI is a threat that needs to be regulated because only they have the answer? I’m shocked.
More seriously, so far I haven’t seen much indication that Mythos is more than Opus with a security focused code analysis harness. That said, the fact it can find these bugs in an automated fashion is the more important takeaway outside of the hype.
I’m curious what the error rate is on the detections, because none of that means much if it is wrong 90% of the time and we are only hearing about the examples that are useful marketing.
>> Anthropic using marketing to convince people their models are more advanced, better built, or that AI is a threat that needs to be regulated because only they have the answer? I’m shocked.
I remember when OpenAI was saying GPT-2 was too dangerous to release.
I remember when there was a guy at Google years a few years ago that was convinced that they had an internal, sentient creature in their labs (I think maybe 4 years ago?)
If I’m not mistaken, after the media cycle, he lost his job for breaking confidentiality.
That was the opposite of marketing, Google really didn’t get how to turn this into a product until ChatGPT happened.
"it can almost like write 2 paragraphs!" "It might be conscious" "this is basically AGI, we had to fire someone who spilled the beans"
I always thought he was fired for making crackpot statements to the press in reference to his professional capacity, and thus creating bad PR and embarrassing spectacle for his employer. Seems like legitimate reasons to me.
An interesting question now is whether he had standard mental health issues, or if he was an early example of AI psychosis or whatever we call people who are falling in love with their AI chatbots because they tell them how smart they are.
Considering Richard Dawkins has recently succumbed to the same delusion it is a reminder that no matter how intelligent someone may otherwise be, we are all human and have certain tendencies and blind spots; anthropomorphizing non-entities being one of those.
Richard Dawkins is 85 to be fair, just like Bernie Sanders is 84 when he made similar comments. The other guy worked on Google's AI safety team where one would assume he'd have a basic grasp of how the technology works.
He also said this [1] a few weeks ago about AI PRs.
> Over the last few months, we have stopped getting AI slop security reports in the #curl project. They're gone.
> Instead we get an ever-increasing amount of really good security reports, almost all done with the help of AI.
> They're submitted in a never-before seen frequency and put us under serious load.
> I hear similar witness reports from fellow maintainers in many other Open Source projects.
> Lots of these good reports are deemed "just bugs" and things we deem not having security properties.
[1]: https://www.linkedin.com/posts/danielstenberg_hackerone-shar...
It may well be that the hype was primarily marketing.
The other alternative is that Curl is simply secure enough that there was far less to find than in other projects.
Given how much money is on the line, it would be gross negligence if anything came publicly out of the CEO's mouth or is otherwise published by the company that's not marketing.
The question is whether they need to massage the results for them to be marketable.
My guess:
Marketing is not intentional.
Evidences: 10 years ago, when I interviewed Baidu AI with Andrew Ng and Dario, Dario is the kind of person is pure-hearted to the point being ideological. Given Dario's successful career so far, that essence has gradually grown into a conviction, and surrounded by a purposely built team which amplifies his ideology.
Humans are very convenient creature, a rare few small fraction of them are no doubt the master of convenience: they morph their mental manifold without a hint of contradiction in their own mental mechanisms.
> Marketing is not intentional
Mythos put Anthropic back into the White House’s good graces. It also branded Anthropic as badass, something their softener image probably needed to win government contracts.
Maybe it wasn’t marketing. But the product’s configuration, and how Anthropic talked about and released it, sure as hell played beautifully. (The timing, while Musk and Altman are distracted with each other, also couldn’t have been better.)
These things are layered. They are great scientists, smart people, etc.
Things change when you’re running a business like Anthropic, especially as the CEO. You have a responsibility to shareholders, and you just need to play the game.
Anthropic chose a great angle: focus on professionals / enterprise, safety, etc. Those can both be done by a genuine desire to make great technology, and for business purposes require you to position yourself in a bit “better” way than reality.
Just look at what their strategy is with Mythos, it’s almost perfection: the “it’s not ready to be released to the public” angle hits all the marks: they care about responsibility / safety, they have “the best” model, and “LLMs are dangerous, but we, as the guardians, can be trusted”. This also helps the industry as a whole with regulation: if they’re being constrained, China will develop even more dangerous models.
This is a result of how smart people treat business, it’s PR perfection, especially given how much the whole industry is talking about it.
(Yes, they fail in other PR areas, but that’s a different discussion)
> Marketing is not intentional.
That's an odd definition of "intentional". Evolution has filtered for people with certain views and the marketing has just emerged from their actions. ... So?
A deadly virus (naturally occurring one let's say) wasn't created intentionally. Evolution selected for it. It's still bad and kills people. Doesn't make it nice because of lack of intention.
I'm not sure if that distinction is important, since what you've described less charitably synonymous with the phrase "Dario is delusional, and has surrounded himself with yes-men, so outlandish marketing gets published as a side effect".
Whether the person doing the marketing was sincere about it or not is immaterial, since marketing is experienced almost entirely by the people consuming it, and not the people communicating it. What matters is if the audience is sincerely concerned by the message, and it's transparently the case that they were sincerely concerned by it.
Curl simply isn't a good data point. It's one of the most picked-over codebases in existence with extensive security testing practices. All the researchers using not-quite-Mythos models have had plenty of time to report bugs up to this point. Daniel may be right that Mythos hasn't been a game changer for curl but the preconditions are different for virtually any other codebase. Perhaps the real marketing here is his own modesty about curl's maturity.
To me, it is a very good data point.
Curl uses all sorts of tools, including AI tools to find bugs. These tools, according to the article found hundreds of bugs including a dozen CVE.
Mythos found one vulnerability. It means the Mythos is just another tool, not the revolution it claims to be.
It is common that when a new tool is introduced that a bunch of bugs are found, with diminishing returns. Mythos finding one vulnerability is consistent to what I would expect for a major update to an existing tool, which Mythos is over existing LLM-based solutions.
The question is how many security vulnerabilities are actually left in the code after all the recent AI attention. Either Mythos is a nothingburger, or it's substantially more powerful but there's nothing left to do. Even a large amount of C can be correct eventually. Curl has the _potential_ to become a good data point maybe 6-12 months from now - if researchers and new tools find many more vulnerabilities then Mythos is proved to be hype. If they don't, then maybe Mythos is overkill for today's curl and its capabilities are better deployed elsewhere (like Firefox, apparently).
I have a hard time believing that Mythos found the only remaining Curl vulnerability. It is possible, but highly improbable.
And it is not overkill, the proof is that it found that vulnerability. It is like saying the new version of some static analyzer with some new rules is "overkill" because it only found only one more bug than the previous version. Deciding whether it is overkill or not is more about context. Using a very expensive model like Mythos for some little used non-critical software is overkill, but for Curl, it absolutely isn't.
If Mythos found loads of vulnerabilities in Firefox but not in Curl, I wouldn't say that's because of Mythos is so good, but rather that with the release of Mythos, they did some testing that could have been done before using the same tools Curl have used.
We will see. As for "testing that could have been done before", Mozilla's posts indicate otherwise. Use of Opus 4.6 led to 22 security-sensitive bugs vs Mythos' 271 (https://blog.mozilla.org/en/privacy-security/ai-security-zer...). They already had the methodology in place when the more powerful model came along (https://hacks.mozilla.org/2026/05/behind-the-scenes-hardenin...):
> Once the end-to-end pipeline is in place, it’s trivial to swap in different models when they become available. Building this pipeline early helped us find a number of serious bugs using publicly-available models, and it also helped us hit the ground running when we had the opportunity to evaluate Claude Mythos Preview. In our experience, model upgrades increase the effectiveness of the entire pipeline: the system gets simultaneously better at finding potential bugs, creating proof-of-concept test cases to demonstrate them, and articulating their pathology and impact.
that makes it a good data point, because it is better able to illustrate the incremental capabilities of Mythos compared to previous tooling
that helps us to understand how much of Mythos is hype and how much is real
We see this exact hypetrain every time a new model is released. Mythos simply hasn't lived up to the "we're all gunna die from the flood of vulnerabilities" hype even slightly. Its slightly better than previous models by all accounts, cool stuff
I've seen literally near word-for-word this exact chain of events multiple times previously
I'm pretty sure mythos is just a new unreleased version of Opus + marketing + a different system prompt.
I suspect so as well.
I've been running my own security scanning software (disclaimer: now starting a company @ zeroquarry.com) for this, and from what I've seen there's a huge value in prompts + adversarial LLM review. Without adversarial review, you get garbage (as this blog points out: 4/5 basically are nonsense) and with a good prompt, you can use almost any "near frontier" model from my experience as long as the prompt helps with the guardrails or the model doesn't protect in such a strict way
They might be biased by the fact that curl is significantly more secure than the average software
Mythos marketing really leans into that "too powerful to be legal" vibe, much like how PS2s were allegedly banned from North Korea because their chips were basically missile-grade.
>It's a good reminder for us all that the competition in this space is rough and lots of more or less subtle marketing is involved.
About as subtle as a personal injury lawyer's billboard
Better Call Dario
A thankfully American reference
Can you expand on this? Do you mean in contrast to the European AI milieu?
No, the personal injury lawyer billboards.
Eh... I think he puts the LLM down for his own ego's sake (as would I!). Curl may, next to the Linux kernel, be one of the most heavily audited codebases in existence. The LLM found something he and thousands of others missed. It's not unimpressive.
I'd go out and say the marketing is not subtle. The hype and fanboys/girls are so in line with the marketing that any level of skepticism is seen a an act of defection, but if you look at the words, hyperbole and volume that is used, there is nothing subtle about it.
It's almost Trump-esque - "this model will change everything forever; we are doomed; we are saved; we will all be fired; we will all be rich", etc
That's a pretty good encapsulation of the parallels between the political and the technological: One necessarily thrives upon the other and are inextricable. This moment is a culmination of all the disenfranchisement the bodypolitik have suffered, looking for any possible means of escape or elevation. AI and Trumpism, for their own respective cohorts, are salvation, on offer by different frontmen but ultimately in service of the same system.
They need the hype to pay off way more than we do. So many of us who still write code directly stand to lose nothing of our capabilities if the marketing claims cannot hold water.
I seem to be totally outside the hype bubble, but I have to suspect there is a lot of imagineering and wild extrapolations in the elss technical hype bubbles. I am curious but no enough to go looking.
>I seem to be totally outside the hype bubble
I'm surprised you say that because it is all over Hacker News. Every single post is co-opted into promoting AI. Try finding a submission with fifty points or more than doesn't have AI or LLM's mentioned somewhere in the comments.
Feel free to retire from the field if you grow tired of seeing its latest developments.
I already have.
That’s not really the point though. I have no doubt AI is useful, I just don’t want to have it shoved in my face every five minutes.
> An amazingly successful marketing stunt for sure.
This. Well done by Antropic.
It even reached the CISO of my small semi-government org in the Netherlands, who slightly panicked at the announced 'tsunami' of vulnerabilities that was coming with Mythos.
Got us some more money and priority with the board, though.
Never waste a good marketing scare.
I don't agree with the "no tsunami in sight": if you don't look at 100+ bugs in Firefox and many more OSS projects, bunch of old unseen-before OpenBSD/Linux RCEs, and a few LPE in just 2 or 3 weeks for Linux itself...
IMO, this does not sound like marketing scare, there is spike of vulnerability disclosures - high quality, low false positives - that can be sensed... It feels like we're speedrunning through few-years worth of high quality bug reports in just a few weeks.
The LPEs were not found with Mythos but with existing, publicly available models.
And also: they did an earlier run with Opus to discover bugs (like segfaults).
In February, Opus discovered a whole bunch of security related bugs, but didn’t exploit them.
Mythos, in turn, was fed these bugs and told to exploit them.
Not saying it’s not impressive, but it was literally told “here are all the places our metal detector says there may be gold, please find gold”.
Anthropic has is quickly destroying customer goodwill by repeatedly pulling the same stunt. Horrible marketing, imho.
It's an entirely different thing to have the company conduct research on LLMs in general being a cybersecurity threat, instead of going "our new model is just too powerful" and shift the discussion to revolve around that. It's slimey.
The bar has become so low lately that no one will care.
org head is smart.
> Not particularly “dangerous”
I'm not sure that follows. As noted, curl was already analyzed to death with every tool available; most software isn't at that level.
But Mythos is not marketed as a tool that can do the same as other tools already available maybe slightly better, but as a revolution.
Sure, but isn't it a verdict on Mythos compared to other models?
If so, it would still follow. "Most software" isn't analyzed as much as curl, by either other tooling or other models, that might well find close to the same as Mythos did. As such, Mythos then isn't especially/particularly dangerous.
I don't think I understand what you mean, the "not particularly dangerous" comment was in relation to the vulnerability that was found right ? Surely they would know what constitutes a lower severity level.
The "not particularly dangerous" is a headline for a section talking about Mythos, not the vulnerability.
Ah okay, that makes a bit more sense. I read it wrong. Then the comment is absolutely fair.
My guess is that it is in category of "you are holding it wrong". Still worth fixing, but requires very specific user input for example. Or very weird scenario. Or in some less used protocol or flag combination.
Curl is currently receiving a record number of high-quality bug/vuln reports (a rather sharp change from the earlier slop inundation), so it’s not like there’s nothing to find. Many or most of these are presumably found by human experts assisted by AI tools, but if Mythos were truly revolutionary, it should be able to find such issues on its own.
https://daniel.haxx.se/blog/2026/04/22/high-quality-chaos/, linked from TFA
Is there a list of infrastructure that has received this kind of focus? Clearly people are looking at the linux kernel, hopefully openssl?
> These tools and the analyses they have done have triggered somewhere between two and three hundred bugfixes merged in curl through-out the recent 8-10 months or so.
If you've just gone through a lengthy analysis of your code with other AI tools, surely it's reasonable not to expect to see hundreds more from a new tool?
It should be possible, unless more bugs are introduced, to eventually get to a state where there are no more bugs in your code.
Process aside, it sounds like Daniel expected to find dozens/hundreds more bugs.
Mythos was kind of hyped as the tool that would discover much more bugs than any currently available tool
> The single confirmed vulnerability is going to end up a severity low CVE planned to get published in sync with our pending next curl release 8.21.0 in late June
My mind still cannot understand the quality and refinement that's gone into cURL. It really is the perfect example of something done so right, that people barely think twice about.
Easy, it shows what is achievable if there is a high bar for quality in every single line of code that gets commited, reviewed and merged, regardless of the programming language.
However in the days of race to bottom, offshoring for penies, and now LLM powered code generation, this is a quality most companies won't care unless there is liability in place.
Curl and SQLite are my favourite examples of properly engineered, rigourously tested _anything_. It's really philosophical - those projects' contribution requirements demand such rigor, and the maintainers stand by that demand. A non-load-bearing document (not project code) is what makes that possible - very reminiscent of Einstein's thought experiments leading to tangible projects such as GPS or Descartes's belief that all problems can be solved through rational thinking.
Some people must be working on training some models exclusively on high quality OSS code base like curl and SQLite without the noise of low quality training data.
I would do that with 100% local models from scratch.
There is always marketing involved and people should be able to put marketing into perspective.
Also curl in this regard is a open source project, relativly small but critical, well known and used everywhere. Besides image libraries, tools like curl or sudo, su, passwd, etc. would also be my first try.
Mythos is still not known at all what it can do. What does it mean from cost and benchmark pov to have a 10 Trillion parameter model?
Nonetheless, the fact that LLMs got significant better in finding this, better than humans, started to happen half a year ago? so at one point we need to address the elefant in the room and state that today you need to do security scanning additional with LLMs. You need to take this serious.
In worst case, use Anthropics marketing to state that its a must now and something changed.
> What does it mean from cost and benchmark pov to have a 10 Trillion parameter model?
To me it means that we've hit the top end of the S-curve with regards to effects of scaling - if the tool isn't remarkably better despite the scale, then we're firmly in diminishing returns territory.
> Mythos is still not known at all what it can do.
And this is very much on purpose my friend. Think about what people already believe it can do though.
> Nonetheless, the fact that LLMs got significant better in finding this, better than humans, started to happen half a year ago?
*rolls eyes* regular static analyzers also have been "better than humans" for decades, being better than a human at a specific mechanical task really doesn't mean much. The interesting new thing is the type of potential "fuzzy bugs" described in the article that LLMs are able to identify (a comment not matching the code it describes, uncommon usage of a 3rd party library, mismatch of code and a protocol it implements, or often just generally weird looking code somebody should have a closer look at... this closes a gap in the traditional debugging toolboxes, but shouldn't replace them)
You don't have to dismantle a comment on a microlevel.
It has been clear for ages that certain type of bugs or issues are better solved from software.
But there was still plenty of things a proper SecOps Person would be able to find with help from tooling which automatic tooling wouldn't find.
Taking a limited amount of resources and focusing on the critical things.
I do think this is gone now. Same with Threat modeling etc.
Putting on my tinfoil-hat: Sooo, the guy who runs the test and delivers the report could just have removed the more interesting bugs and delivered those to any three letter agency?
curl's source is public so what would be the gain in the rigmarole? Now if the prompt was "create a patch that inserts a zero-day while fixing a bug" that would be impressive.
Swival found many more vulnerabilities without Mythos https://github.com/swival/security-audits
I don't know about Mythos but in recent weeks I've noticed Opus is constantly failing to fix things in tsz[0] vs GPT 5.5 can easily churn out fixes that are solid and pass tests. I've stopped paying for Claude for now and all my money is going to OpenAI at the moment. Either Opus is massively nerfed or GPT 5.5 is really head and shoulder higher in terms of very difficult tasks. The last percent of conformance tests in tsz are really really difficult and I've seen Opus bailing again and again. So annoying to waste time and tokens to finally get "this is too involved" or "this requires a multi-week sprint to fix".
[0] https://tsz.dev
The new Opus feels like a step backwards. More expensive, thinks more, and it does not get the job done.
From a user’s perspective 4.7 is a downgrade compared to 4.6 . It’s intended to give Anthropic more control about their compute resources and profitability:
https://news.ycombinator.com/item?id=48072916
Having never used Claude and only Codex, does Claude actually say “this is too involved” as a response to a prompt?
Yes it does. Usually after hours of working and not getting results
I am curious, what kind of work do you use Claude for that sometimes requires hours of working. In my case, I have never seen it go off for more than 10 mins and even that is very rare.
debugging code. I had some issue so I create a plan to root cause that would run the code, change some functions or variables and run again until we get a confirmed answer.
I just work up to that very workflow this morning. I ran last night and finished at around 3am with ~200k tokens spent. Fixed the issue and created a follow up doc for things that it could not verify.
https://github.com/mohsen1/tsz
I'm looking forward to trying Mythos run against my 5000-line, instant-finality, quantum-resistant blockchain project and decentralized exchange (an additional 5000 lines). I already ran all the models up to Opus 4.6 and they couldn't find anything.
I routinely used to compile C programs on other compilers to find defects that one or another didn't find. Compiling on Windows vs Linux. You could summarize / minimize it down to compiling it with warning as errors etc but you'd be missing the point.
The point wasn't actual cross-platform portability even though that was a nice side effect. It was to flush out all the weird edge cases.
Edges like security flaws. Buffer overflows are usually platform specific. There are plenty of other ways to find these issues but simply recompiling for a different platform surfaces all sorts of issues.
> The source code consists of 660,000 words, which is 12% more words than the entire English edition of the novel War and Piece.
Typo, or is there a spoof I should go read?
Perhaps he was dictating.
Does it say anything else? Just 'Aaaarggghhhh'?
Doubt it considering that Daniel Stenberg is Swedish. English dictation when you speak English as a second language with an accent is quite annoying.
Voice input works really well for people speaking English with a Swedish accent. I think the accent of most educated Swedes is mostly a case of prosody. For sure there are some sounds we say slightly differently than native English speakers. We often have some trouble with /s/ and /z/, but I don't know, "war and peace", I think that's easily understood.
Source: voice typing this with Swedish vocal chords, and only had to correct "different lives" to "differently", and add /[^\w\s]/.
Android voice input works with kids using both English and native words, here in India. The country runs schools in 25+ primary languages, each with dialects, so a TV/phone with voice input is more marvelous than the nitpicks discussed here.
I understand completely. You don't want to know what the machine produced, when I asked it for "a new display".
War and Peace is about 590,000 words. Tiny compared to the full Harry Potter collection (about 1 million words over the 7 books), but long for a single book.
They're referring to the typo in the title, "Piece" vs "Peace".
I also thought they were contending the word count before noticing. Even remarked how I find this a weird metric, given that code is not prose [0], but then I deleted that once I picked up on what's going on.
[0] comparing the output of `wc -w` with the word counts of books I'm reasonably sure will be super off
edit: ran a calc, substituting out symbols (but not underscores), digits, and comments yields a 390K word count compared to the 660K cited. not excluding the comments yields 600K, so more than a third of all words in the sources are comments.
The ten main Malazan books are 3.3 million words, apparently. No wonder it took me such a long time to get through them.
It's a shame he seems to reject the idea of actually diving in and using these tools interactively:
> It’s not that I would have a lot of time to explore lots of different prompts and doing deep dive adventures anyway.
His expertise I think would elevate the results quite a bit. Although if he never uses LLMs, which it reads like he doesn't, I guess it might backfire just as well. Prompting style (still?) does matter after all, certainly in my experience anyways.
He states in the article that they use LLMs for this purpose and find them extremely useful.
Which can be true without this also being true:
> using these tools interactively
I did read the article. It seems to me they're using LLMs in a prepared manner instead, as mere scanners that produce reports.
He posts about his use of language models a lot on Mastodon[0]. He does lots with language models, but doesn't buy all the way into the hype. I'd say he's one of. most reasonable & balanced voices on the subject of AI use in software today. Happy to use the technology, more than willing to push back on marketing bs.
[0] https://mastodon.social/@bagder
Won my bet "voted 10 [vulnerabilities] but in retrospect as you are familiar with Claude and such tooling if you already used any of recent model to done some kind of security review then I'd drop to 1 or even 0." https://mastodon.pirateparty.be/@utopiah/116537456780283420