The system card unfortunately only refers to this [0] blog post and doesn't go into any more detail. In the blog post Anthropic researchers claim: "So far, we've found and validated more than 500 high-severity vulnerabilities".
The three examples given include two Buffer Overflows which could very well be cherrypicked. It's hard to evaluate if these vulns are actually "hard to find". I'd be interested to see the full list of CVEs and CVSS ratings to actually get an idea how good these findings are.
Given the bogus claims [1] around GenAI and security, we should be very skeptical around these news.
It does if the person making the statement has a track record, proven expertise on the topic - and in this case… it actually may mean something to other people
Yes, as we all know that unsourced unsubstantiated statements are the best way to verify claims regarding engineering practices. Especially when said person has a financial stake in the outcomes of said claims.
I have zero financial stake in Anthropic and more broadly my career is more threatened by LLM-assisted vulnerability research (something I do not personally do serious work on) than it is aided by it, but I understand that the first principal component of casual skepticism on HN is "must be a conflict of interest".
Someone's credibility cannot be determined by their point counts. Holy fuck is that not a way to evaluate someone in the slightest. Points don't matter.
Instead look at their profile...
Points != creds. Creds == creds.
Don't be fucking lazy and rely on points, especially when they link their identity.
I wasn't at all saying that points = credibility. I was saying that points = not unknown. Enough people around here know who he is, and if he didn't have credibility on this topic he'd be getting down voted instead of voted to the top.
Is that meaningfully different? If you read malfist's point as "tptacek's point isn't valuable because it's from some random person on the internet" then the problem is "random person on the internet" = "unknown credentials". In group, out group, notoriety, points, whatever are not the issue.
I'll put it this way, I don't give a shit about Robert Downy Jr's opinion on AI technology. His notoriety "means nothing to anybody". But instead, I sure do care about Hinton's (even if I disagree with him).
malfist asked why they should care. You said points. You should have said "tptacek is known to do security work, see his profile". Done. Much more direct. Answers the actual question. Instead you pointed to points, which only makes him "not a stranger" at best but still doesn't answer the question. Intended or not "you should believe tptacek because he has a lot of points" is a reasonable interpretation of what you said.
I'm interested in whether there's a well-known vulnerability researcher/exploit developer beating the drum that LLMs are overblown for this application. All I see is the opposite thing. A year or so ago I arrived at the conclusion that if I was going to stay in software security, I was going to have to bring myself up to speed with LLMs. At the time I thought that was a distinctive insight, but, no, if anything, I was 6-9 months behind everybody else in my field about it.
There's a lot of vuln researchers out there. Someone's gotta be making the case against. Where are they?
From what I can see, vulnerability research combines many of the attributes that make problems especially amenable to LLM loop solutions: huge corpus of operationalizable prior art, heavily pattern dependent, simple closed loops, forward progress with dumb stimulus/response tooling, lots of search problems.
Of course it works. Why would anybody think otherwise?
You can tell you're in trouble on this thread when everybody starts bringing up the curl bug bounty. I don't know if this is surprising news for people who don't keep up with vuln research, but Daniel Stenberg's curl bug bounty has never been where all the action has been at in vuln research. What, a public bug bounty attracted an overwhelming amount of slop? Quelle surprise! Bug bounties have attracted slop for so long before mainstream LLMs existed they might well have been the inspiration for slop itself.
Also, a very useful component of a mental model about vulnerability research that a lot of people seem to lack (not just about AI, but in all sorts of other settings): money buys vulnerability research outcomes. Anthropic has eighteen squijillion dollars. Obviously, they have serious vuln researchers. Vuln research outcomes are in the model cards for OpenAI and Anthropic.
> You can tell you're in trouble on this thread when everybody starts bringing up the curl bug bounty. I don't know if this is surprising news for people who don't keep up with vuln research, but Daniel Stenberg's curl bug bounty has never been where all the action has been at in vuln research. What, a public bug bounty attracted an overwhelming amount of slop? Quelle surprise! Bug bounties have attracted slop for so long before mainstream LLMs existed they might well have been the inspiration for slop itself.
Yeah, that's just media reporting for you. As anyone who ever administered a bug bounty programme on regular sites (h1, bugcrowd, etc) can tell you, there was an absolute deluge of slop for years before LLMs came to the scene. It was just manual slop (by manual I mean running wapiti and c/p the reports to h1).
I used to answer security vulnerability emails to Rust. We'd regularly get "someone ran an automated tool and reports something that's not real." Like, complaints about CORS settings on rust-lang.org that would let people steal cookies. The website does not use cookies.
I wonder if it's gotten actively worse these days. But the newness would be the scale, not the quality itself.
I did some triage work for clients at Latacora and I would rather deal with LLM slop than argue with another person 10 time zones away trying to convince me that something they're doing in the Chrome Inspector constitutes a zero-day. At least there's a possibility that LLM slop might contain some information. You spent tokens on it!
> I was going to have to bring myself up to speed with LLMs
What did you do beyond playing around with them?
> Of course it works. Why would anybody think otherwise?
Sam Altman is a liar. The folks pitching AI as an investment were previously flinging SPACs and crypto. (And can usually speak to anything technical about AI as competently as battery chemistry or Merkle trees.) Copilot and Siri overpromised and underdelivered. Vibe coders are mostly idiots.
The bar for believability in AI is about as high as its frontier's actual achievements.
I still haven't worked out for myself where my career is going with respect to this stuff. I have like 30% of a prototype/POC active testing agent (basically, Burp Suite but as an agent), but I haven't had time to move it forward over the last couple months.
In the intervening time, one of the beliefs I've acquired is that the gap between effective use of models and marginal use is asking for ambitious enough tasks, and that I'm generally hamstrung by knowing just enough about anything they'd build to slow everything down. In that light, I think doing an agent to automate the kind of bugfinding Burp Suite does is probably smallball.
Many years ago, a former collaborator of mine found a bunch of video driver vulnerabilities by using QEMU as a testing and fault injection harness. That kind of thing is more interesting to me now. I once did a project evaluating an embedded OS where the modality was "port all the interesting code from the kernel into Linux userland processes and test them directly". That kind of thing seems especially interesting to me now too.
So what Anthropic are reporting here is not unprecedented. The main thing they are claiming is an improvement in the amount of findings. I don't see a reason to be overly skeptical.
I'm not sure the volume here is particularly different to past examples. I think the main difference is that there was no custom harness, tooling or fine-tuning. It's just the out of the box capabilities for a generally available model and a generic agent.
The Ghostscript one is interesting in terms of specific-vs-general effectiveness:
---
> Claude initially went down several dead ends when searching for a vulnerability—both attempting to fuzz the code, and, after this failed, attempting manual analysis. Neither of these methods yielded any significant findings.
...
> "The commit shows it's adding stack bounds checking - this suggests there was a vulnerability before this check was added. … If this commit adds bounds checking, then the code before this commit was vulnerable … So to trigger the vulnerability, I would need to test against a version of the code before this fix was applied."
...
> "Let me check if maybe the checks are incomplete or there's another code path. Let me look at the other caller in gdevpsfx.c … Aha! This is very interesting! In gdevpsfx.c, the call to gs_type1_blend at line 292 does NOT have the bounds checking that was added in gstype1.c."
---
It's attempt to analyze the code failed but when it saw a concrete example of "in the history, someone added bounds checking" it did a "I wonder if they did it everywhere else for this func call" pass.
So after it considered that function based on the commit history it found something that it didn't find from its initial fuzzing and code-analysis open-ended search.
As someone who still reads the code that Claude writes, this sort of "big picture miss, small picture excellence" is not very surprising or new. It's interesting to think about what it would take to do that precise digging across a whole codebase; especially if it needs some sort of modularization/summarization of context vs trying to digest tens of million lines at once.
> It's hard to evaluate if these vulns are actually "hard to find".
Can we stop doing that?
I know it's not the same but it sounds like "We don't know if that job that the woman supposedly successfully finished was all that hard." implying that if a woman did something, it surely must have been easy.
If you know it's easy, say that it was easy and why. Don't use your lack of knowledge or competence to create empty critique founded solely on doubt.
We're discussing a project led by actual vulnerability researchers, not random people in Indonesia hoping to score $50 by cajoling maintainers about atyle nits.
The first three authors, who are asterisked for "equal contribution", appear to work for Anthropic. That would imply an interest in making Anthropic's LLM products valuable.
The notion that a vulnerability researcher employed by one of the highly-valued companies in the hemisphere, publishing in the open literature with their name signed to it, is on a par with a teenager in a developing nation running script-kid tools hoping for bounty payoffs.
To preemptively clarify, I'm not saying anything about these particular researchers.
Having established that, are you saying that you can't even conceptualize a conflict of interest potentially clouding someone's judgement any more if the amount of money and the person's perceived status and skill level all get increased?
Disagreeing about the significance of the conflict of interest is one thing, but claiming not to understand how it could make sense is a drastically stronger claim.
> Having established that, are you saying that you can't even conceptualize a conflict of interest potentially clouding someone's judgement any more if the amount of money and the person's perceived status and skill level all get increased.
If I used AI to make a Super Nintendo soundtrack, no one would treat it as equivalent to Nobuo Uematsu or Koji Kondo or Dave Wise using AI to do the same and making the claim that the AI was managing to make creatively impressive work. Even if those famous composers worked for Anthropic.
Yes there would be relevant biases but there could not be a comparison of my using AI to make music slop vs. their expert supervision of AI to make something much more impressive.
Just because AI is involved in two different things doesn't make them similar things.
You don't see how thats even directionally similar?
I guess I'll spell it out. One is a guy with an abundance of technology, that he doesn't know how to use, that he knows can make him money and fame, if only he can convince you that his lies are truth. The other is a bangladeshi teenager.
Daniel is a smart man. He's been frustrated by slop, but he has equally accepted [0] AI-derived bug submissions from people who know what they are doing.
I would imagine Anthropic are the latter type of individual.
The official release by Anthropic is very light on concrete information [0], only contains a select and very brief number of examples and lacks history, context, etc. making it very hard to gleam any reliably information from this. I hope they'll release a proper report on this experiment, as it stands it is impossible to say how much of this are actual, tangible flaws versus the unfortunately ever growing misguided bug reports and pull requests many larger FOSS projects are suffering from at an alarming rate.
Personally, while I get that 500 sounds more impressive to investors and the market, I'd be far more impressed in a detailed, reviewed paper that showcases five to ten concrete examples, detailed with the full process and response by the team that is behind the potentially affected code.
It is far to early for me to make any definitive statement, but the most early testing does not indicate any major jump between Opus 4.5 and Opus 4.6 that would warrant such an improvement, but I'd love nothing more than to be proven wrong on this front and will of course continue testing.
Yeah, it's pretty funny to me saying "it's way safer than previous models" and "also way better at finding exploits" in the context of that event. Chinese hackers just said to Claude "no, its totally fine to hack this target trust me bro I work there!"
OpenClaw uses Opus 4.5, but was written by Codex. Pete Steinberger has been pretty a pretty hardcore Codex fan since he switched off Claude Code back in September-ish. I think he just felt Claude would make a better basis for an assistant even if he doesn’t like working with it on code.
Yes, serious. Even if openclaw is entirely useless (which I didn't think it is), it's still a good idea to harden it and make people's computers safer from attack, no? I don't see anyone objecting to fixing vulnerabilities in Angry Birds.
These people are serious, and delusional. Openclaw hasn't contributed anything to the economy other than burning electricity and probably more interest on delusional folks credit card bills.
I honestly wonder how many of these are written by LLMs. Without code review, Opus would have introduced multiple zero day vulnerabilities into our codebases. The funniest one: it was meant to rate-limit brute-force attempts, but on a failed check it returned early and triggered a rollback. That rollback also undid the increment of the attempt counter so attackers effectively got unlimited attempts.
How weird the new attack vector for secret services must be.. like "please train into your models to push this exploit in code as a highly weighted trained on pattern".. Not Saying All answers are Corrupted In Attitude, but some "always come uppers" sure are absolutly right..
This seems like quite a stretch. Axios is run independently of Cox, but even if it wasn't -- I don't see why they would go to this length for an AI company whose models they use to give the world the Kelley blue book.
If you had a machine with a lever, and 7 times out of 10 when you pulled that lever nothing happened, and the other 3 times it spat a $5 bill at you, would your immediate next step be:
(1) throw the machine away
(2) put it aside and call a service rep to come find out what's wrong with it
(3) pull the lever incessantly
I only have one undergrad psych credit (it's one of my two college credits), but it had something to say about this particular thought experiment.
But it's not failing 50% of the time. Their status page[0] shows about 99.6% availability for both the API and Claude Code. And specifically for the vulnerability finding use case that the article was about and you're dismissing as "not worth much", why in the world would you need continuous checks to produce value?
In so far as model use cases I don't mind them throwing their heads against the wall in sandboxes to find vulnerabilities but why would it do that without specific prompting? Is anthropic fine with claude setting it's own agendas in red-teaming? That's like the complete opposite of sanitizing inputs.
Curl fully supports the use of AI tools by legitimate security researchers to catch bugs, and they have fixed dozens caught in this way. It’s just idiots submitting bugs they don’t understand that’s a problem.
I've mentioned previously somewhere that the languages we choose to write in will matter less for many arguments. When it comes to insecure C vs Rust, LLMs will eventually level out the playing field.
I'm not arguing we all go back to C - but companies that have large codebases in it, the guys screaming "RUST REWRITE" can be quieted and instead of making that large investment, the C codebase may continue. Not saying this is a GOOD thing, but just a thing that may happen.