Don't get me wrong, it's interesting. But there is no technical discussion as to how they did it. It's simply: we did it and Mythos and Codex didn't.
It's good to know that it's possible, but I'd have already expected it. Put a base model versus a base model + harness + whatever else, and yea, if you do it right then you have a better system to find vulnerabilities.
> We then ran AISLE's autonomous AI system against curl.
They don't even mention what models the use under the hood. It wouldn't surprise me if they are from Anthropic and OpenAI.
The homepage says something about AI guided fuzzing based on libfuzzer or AFL. Looks like they have the LLMs identify a bunch of interesting functions to test, generate some test harnesses, and then sort through the fuzzer findings at a high level, which sounds like a pretty good idea.
Fuzzing or having the LLM sort through where might be most useful to fuzz & sorting the results? Neither seem particularly compute intensive to me, fuzzing is a pretty standard step and having the LLM read through to find the most interesting areas to fuzz sounds a lot more efficient than leaving the whole task to the LLM.
Setting a swarm of agents loose for hours to look for software vulnerabilities is far more compute-expensive than fuzzing. The industry has never thrown this kind of compute resources at pure fuzzing, in part because you can't get much VC money for that.
Thanks for figuring that out. Sort of sounds like AI programming programs to find vulnerabilities, of which fuzzing is one of the proven techniques to do it.
It wouldn’t surprise me if AISLE uses many different providers’ models, and what’s holding back OpenAI and Anthropic is only using first-party models. Just because OpenAI and Anthropic have arguably the strongest models overall doesn’t mean their models are the strongest at finding any given class of vulnerability or lead to follow.
Maybe the model doesn’t matter, maybe you just need something minimally intelligent to seed the fuzzer, generate a test case, and rinse and repeat when the fuzzer gets stuck.
The tool basically had to chain two exploits together to reach this. It also came up with a patch to fix which was fairly sensible (but I ended up editing it further for clarity).
Since AISLE reported 29 issues but only 6 warranted a CVE, and all the found CVEs were "low" severity, this makes me wonder if AISLE simply is tuned for a higher false positive rate than the anthropic and openai tools (which may have found the same 6 issues and decided not to report them)
i don't think this is correct. if you look at this article by the curl founder daniel stenberg (https://daniel.haxx.se/blog/2026/05/11/mythos-finds-a-curl-v...), he talks about how he previously ran Mythos on curl and that it found 5 issues: 1 turned out to be a low severity CVE, 3 were false positives, and 1 just a bug. So a) Mythos detects low severity CVEs too, and b) it is fairly noisy
Good marketing and definitive proof that local (read: on-prem & air-gapped) models with correct context and tools are good enough to perform on par and above SOTA cloud hosted solutions.
We have seen this point many times before with different technologies. The first computers at university were big and expensive, same as this machine. Give it a few years and this functionality will be a commodity.
OpenAI and Anthropic have both been studying CURL for a while though. Anything they found was already fixed.
If you want to compare you need to start with something that none of studied. Somebody please take the source to a 2023 release of CURL (It shouldn't be hard to find one) - before all the current AI craze, and run all the tools on them to see what they find. Only then can we compare numbers. (and even then severity may come into place - all 6 are rated low impact)
I think you might be misunderstanding this? This is, from my understanding, what went down:
1. curl was scanned by many different things, including AISLE, and many bugs were fixed <- all this was in the past
2. curl a week ago was scanned again my Mythos and Codex Security, and both of them said: 0 issues found
3. the same curl was scanned by AISLE a day later, resulting in ~29 reports (based on the blog post and mastodon posts from Daniel Stenberg)
4. of these 29, 6 cleared the bar and got CVEs in curl
5. these 6 CVEs were just announced as fixed in curl 8.22.0 today, together with 4 more CVEs that were detected by other people prior to point 2. of this list
so imho it was head-to-head, the very same codebase => it's a legit comparison
"How many total vulnerabilities can your tool alone identify?" and "How many unique vulnerabilities can your tool identify?" are both valid comparisons to make IMO.
I guess this would also require models trained on pre-2023 data - or not trained on later curl code, changelogs, blog posts discussing curl security fixes, etc.
i don't think this is doable fairly. as they say in the blog post, the only fair way to is to look for new, previously undiscovered zero-days, otherwise you always risk the model has in some way been trained on the vulnerabilities. looking for legit new stuff is the only way to prevent leakage (even accidental one)
Wow, this announcement is good content marketing.
Don't get me wrong, it's interesting. But there is no technical discussion as to how they did it. It's simply: we did it and Mythos and Codex didn't.
It's good to know that it's possible, but I'd have already expected it. Put a base model versus a base model + harness + whatever else, and yea, if you do it right then you have a better system to find vulnerabilities.
> We then ran AISLE's autonomous AI system against curl.
They don't even mention what models the use under the hood. It wouldn't surprise me if they are from Anthropic and OpenAI.
The homepage says something about AI guided fuzzing based on libfuzzer or AFL. Looks like they have the LLMs identify a bunch of interesting functions to test, generate some test harnesses, and then sort through the fuzzer findings at a high level, which sounds like a pretty good idea.
Also sounds incredibly compute intensive.
Fuzzing or having the LLM sort through where might be most useful to fuzz & sorting the results? Neither seem particularly compute intensive to me, fuzzing is a pretty standard step and having the LLM read through to find the most interesting areas to fuzz sounds a lot more efficient than leaving the whole task to the LLM.
Setting a swarm of agents loose for hours to look for software vulnerabilities is far more compute-expensive than fuzzing. The industry has never thrown this kind of compute resources at pure fuzzing, in part because you can't get much VC money for that.
Thanks for figuring that out. Sort of sounds like AI programming programs to find vulnerabilities, of which fuzzing is one of the proven techniques to do it.
It defaults to gpt5.4 nano
https://github.com/weareaisle/nano-analyzer/blob/main/scan.p...
A repo named "nano-analyzer" unsurprisingly uses gpt5.4 nano. I doubt their "pay them money" version uses nano.
Their system can run with various models, they go into more details in this article.
https://aisle.com/blog/system-over-model-zero-day-discovery-...
It wouldn’t surprise me if AISLE uses many different providers’ models, and what’s holding back OpenAI and Anthropic is only using first-party models. Just because OpenAI and Anthropic have arguably the strongest models overall doesn’t mean their models are the strongest at finding any given class of vulnerability or lead to follow.
> what models the use under the hood
Presumably their own, wouldn’t they?
Default to gpt 5.4 nano
https://github.com/weareaisle/nano-analyzer/blob/main/scan.p...
You mean their own trained models, or do you think it's an open source model that they fine-tuned? If they use their own, I'd guess it's the latter.
Maybe the model doesn’t matter, maybe you just need something minimally intelligent to seed the fuzzer, generate a test case, and rinse and repeat when the fuzzer gets stuck.
We had a few AISLE-generated security reports, and the signal to noise was reasonably good.
The most notable bug/exploit their scanner found was: https://gitlab.com/nbdkit/libnbd/-/commit/e50bbd2681117c2dd8...
The tool basically had to chain two exploits together to reach this. It also came up with a patch to fix which was fairly sensible (but I ended up editing it further for clarity).
Since AISLE reported 29 issues but only 6 warranted a CVE, and all the found CVEs were "low" severity, this makes me wonder if AISLE simply is tuned for a higher false positive rate than the anthropic and openai tools (which may have found the same 6 issues and decided not to report them)
As far as I understand it, the other efforts have not reported most of their findings to upstream developers, focusing on critical findings only.
This is understandable because upstream interactions at scale are difficult.
In the case of big projects like curl the collaboration seems a bit more direct.
i don't think this is correct. if you look at this article by the curl founder daniel stenberg (https://daniel.haxx.se/blog/2026/05/11/mythos-finds-a-curl-v...), he talks about how he previously ran Mythos on curl and that it found 5 issues: 1 turned out to be a low severity CVE, 3 were false positives, and 1 just a bug. So a) Mythos detects low severity CVEs too, and b) it is fairly noisy
Curl seems to becoming one of the favourite things to demo AI finding vulns.
Curl is going to end up incredibly secure.
Thank goodness because curl is a load bearing structure to the backend of the internet.
Good marketing and definitive proof that local (read: on-prem & air-gapped) models with correct context and tools are good enough to perform on par and above SOTA cloud hosted solutions.
We have seen this point many times before with different technologies. The first computers at university were big and expensive, same as this machine. Give it a few years and this functionality will be a commodity.
That's bragging rights correctly earned, i think! As marketing-y as this post is, definitely something to keep an eye on.
This is an ad. I didn't learn anything from reading it.
Marketing Slop
OpenAI and Anthropic have both been studying CURL for a while though. Anything they found was already fixed.
If you want to compare you need to start with something that none of studied. Somebody please take the source to a 2023 release of CURL (It shouldn't be hard to find one) - before all the current AI craze, and run all the tools on them to see what they find. Only then can we compare numbers. (and even then severity may come into place - all 6 are rated low impact)
I think you might be misunderstanding this? This is, from my understanding, what went down:
1. curl was scanned by many different things, including AISLE, and many bugs were fixed <- all this was in the past 2. curl a week ago was scanned again my Mythos and Codex Security, and both of them said: 0 issues found 3. the same curl was scanned by AISLE a day later, resulting in ~29 reports (based on the blog post and mastodon posts from Daniel Stenberg) 4. of these 29, 6 cleared the bar and got CVEs in curl 5. these 6 CVEs were just announced as fixed in curl 8.22.0 today, together with 4 more CVEs that were detected by other people prior to point 2. of this list
so imho it was head-to-head, the very same codebase => it's a legit comparison
"How many total vulnerabilities can your tool alone identify?" and "How many unique vulnerabilities can your tool identify?" are both valid comparisons to make IMO.
I guess this would also require models trained on pre-2023 data - or not trained on later curl code, changelogs, blog posts discussing curl security fixes, etc.
i don't think this is doable fairly. as they say in the blog post, the only fair way to is to look for new, previously undiscovered zero-days, otherwise you always risk the model has in some way been trained on the vulnerabilities. looking for legit new stuff is the only way to prevent leakage (even accidental one)