Is the company selling a magical code generation machine really asking how it can backport a fix to an older version of the same software?
Is this for real?
I wouldn't expect the PHP team to provide patches for versions for which security support has ended. But this seems like the kind of task that AI is especially well suited for-- rework the patch for a release that is no longer supported but still widely used. It would've been nice for people who have much older sites to be able to patch their PHP rather than have to upgrade PHP as well to get the fix.
Check it out: PatchWave (https://patchwave.ai)
(Disclosure - I am a co-founder)
> How did we find this?
> No special harness, or access to models that weren't generally available to everyone were used in the initial bug discovery. I focused on something I had a background in, knew its vulnerability history well, and held a theory that specific bugs lurk within the codebase. This shows that "know your target" is still as applicable as ever, even in the LLM dominated VR age. No model is simply capable of unearthing "all bugs" with simple prompts. At least for now.
> There are some practical research implications here. LLMs may have ingested public historical vuln data, but in their default variant analysis tends to move from the newest and most prominent bugs toward nearby older code. They do not reliably begin with obscure twenty-year-old exploitation primitives. Keeping your own mental catalog of a target's vulnerability history lets you prompt that direction explicitly.
> To our knowledge, frontier AI found these vulnerabilities before attackers did
I don't know who/what Calif is and there's no link to them, so if I were only reading this article I would definitely get the impression that "AI did all the work here".
It’s just gross and manipulative. You have no moat, get over it
pAtCh ThE pLaNeT today, tomorrow it's creepy rent-seeking behavior
I feel like taking every action these companies take as part of some larger strategy is kind of conspiracy-minded.
I think the effort is misplaced. We should be exploring ways to make our systems more resilient and I would expect those companies* to focus on that front. Vulnerability research is important, and we should aim to remove as many bugs as we can from our codebases but the pace of software development is increasing and so is the introduction of new bugs.
* I would also expect an increase on relevant research regarding novel paradigms of securing systems.
Where's the conspiracy. There is no moat. The technology is easily reproducible. Deepseek unwound an entire business model yesterday.
But they could be, that's why it's a conspiracy theory.
Other parts of the company want to support the public.
The first part of the company backs up the second in internal politics.
Do you consider this a conspiracy? Just seems like “good business” to me. It doesn’t seem as clean as you want to make it seem, though.
> 19–24 August 2026: Wiz Research independently collides with the same bug family using Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol, reporting its variants on 24 August.
I again ask how they are lying? I get if people hate these companies, but blatantly disregarding what is said helps nobody.
The same view point is used for racism, dei, and the rest: it doesn't matter if the poor beggar on the corner hates white people. They don't have any means or methods to act on it.
CEOs of all big corporations are constantly lying and misleading the public, investors and shareholders. That is a crime. A shame that it is not enforced anymore.
But of course there would be no marketing.
[1] From whom you stole everything in the first place.
Most people license out the software under specific requirements and prohibitions. BSD requires attribution. GPL variants require adherence to copyleft when distributing derivative works. A little under half of all github repos don't even have a license, which means that though the source is published, no one has any license under which to copy it or derive new works from it.
So far as I have seen, whether aggregating and munging these works into abstract LLM weights constitutes fair use is still legally unclear. If not deemed fair use, it is almost certainly a massive violation of copyright.
If you don't want your code to be used as possible inspiration, then you are not actually for open source, and are using the wrong license. The licenses are legal bindings, not moral bindings.
"Don't use any concept found within this code" would be a license that doesn't really fit with any current copyright law, and it wouldn't allow any software to exist, except that made in the 70's.
Or at least just childishly naive for failing to notice that beforehand.
That is, they knowingly and willingly broke the law by stealing an unprecedented volume of proprietary, copyrighted materials.
Books: Bartz v. Anthropic
Films and TV: Disney Enterprises Inc. v. Midjourney Inc.
Music: Concord Music Group, Inc. v. Anthropic
Photography: Getty Images v Stability AI
Journalism: Postmedia Network Inc., Torstar Corporation, The Globe and Mail Inc., and The Canadian Press v. Anthropic PBC
[0] https://en.wikipedia.org/wiki/Grand_Upright_Music%2C_Ltd._v.....
The point of the open source I work on is to contribute to a commons, not to have it stolen so that it can be used to destroy those commons, used for the profit of techbros, and used to justify the murder of children.
To be further pedantic, is that really stealing? They are disregarding copyright but was anything taken from anybody? I thought that was the old argument for piracy, ie, nobody was deprived of anything.
This works for established profitable brands that want to build / maintain a good public image.
This isn’t the marketing needed by an unprofitable startup that wants to show hyper growth before its IPO. They need the “use our product or get pwned” kind of marketing.
I'm sure they named it after the "Hack the planet" slogan from the Hackers movie.
Some people are acting like this is only a problem because labs are releasing models that can find and exploit these vulnerabilities in the first place. This is a nonsense claim. Humans have been discovering vulnerabilities and privately weaponizing exploits just fine for decades. We are collectively raising the floor for security.
Just give us the damn code to verify things are working since so many have to use stopgap solutions and want to verify them
So how many tokens can I put you down for.
Then, someone releases a pick that works for many more locks humans couldn't pick.
And says: You can't stop the progress. Now pay us to upgrade your locks (no-no, they won't be doctored to fit our next-gen picks, trust us!)
Response: "What's the prompt?"
Yep. I might take minimum wage interns and make million dollar software(or million dollars in 2023), but they often run into issues that a 5 minute review with me solves.
I do wonder if AI is going to eliminate prompt engineering... It's basically all I have left. I feel like I have magic words.
What number in line do you expect to be when asking the same prompt.
For example, why wouldn't OpenAI ask that first long before you get a chance at the model?
If they asked their unreleased bot to get a million dollars and it suggests betting on Kalshi, that would be hilarious.
Why wouldn't someone with access long before you ask first? For example a model provider like hugging face, or a nation state looking to interrupt some other nations economy?
The problem with any simple question is you're exceptionally unlikely to be the first person to explore the problem space and pick the lowest hanging fruit. It's not impossible, it's just improbable.
--
Lastly why wouldn't OpenAI ask how to make a million dollars a million times? Or if it's cheap enough for them, how to make $1 a trillion times. There is a reason people talk about AI risks, and one particular risk being that the AI companies no longer need us pesky humans to think for them. They hoover up all the GDP and leave the rest of us to fight for scraps.
How much time will be left in the Defender’s Window between now and an organization getting Daybreak Access?
OpenAI consistently doesn't know what they're talking about.
I thought that logic had been widely discredited in the security community long ago.
> No special harness, or access to models that weren't generally available to everyone were used in the initial bug discovery. I focused on something I had a background in, knew its vulnerability history well, and held a theory that specific bugs lurk within the codebase. This shows that "know your target" is still as applicable as ever, even in the LLM dominated VR age. No model is simply capable of unearthing "all bugs" with simple prompts.
They used a public model as a research assistant and a code assistant. They guided it with expert deep knowledge of the project. OpenAI did not lead this, ChatGPT was not instrumental in this.
It is perhaps at best to say OpenAI made it possible to find the bug more quickly. And on the double edged sword, also made it easier for it to be exploited days ahead of release. So I agree with the original commentator, did they on the balance actually do anything good here?
totally. I spent yesterday trying to get an AI to make a test case for a very specific bug I had. It kept telling me that it had it. When I could clearly trace thru the code and see it did not. After about 5-10 rounds of that I just wrote the test myself. But at least AI helped with all of the dumb cruft that went around it.
But if I did not know the domain I may have easily just said 'yep looks good'. AI on the other hand saved me at least 3-4 hours of writing all of the scaffolding that made it easier for me to check. In the end the bug had nothing to do with the test. AI did not find that. I went back point by point of what changed in the code and found where I had messed something 2 weeks ago.
A few months ago it seemed like it would not do this. Now in the past week I have come across 3 different ones that did something similar. Just 'stuck' in a local maxima and getting the AI to go around it is a pain. This is probably me and the way I am using it.
Is this better than I had a couple of years ago? I am not sure. Seems different at least.
https://deadneurons.substack.com/p/php2shell-explained-php2s...
> 1 August 2026 [...] Ryan Austin finds the RCE paths, tweets a cryptographic hash on 5 August, and submits the bug and working RCE proof-of-concept to PHP between 11 August and 14 August.
> 19–24 August 2026: Wiz Research independently collides with the same bug family using Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol, reporting its variants on 24 August.
I get some skepticism about marketing copy, but your comment is blatantly wrong and needlessly rude.
This is absurd, and it shows that you have absolutely 0 idea of how any of this works. Finding a bug and delivering a working RCE 11 days later is absolutely not "sitting on it". This is how this works. You need time to test, exploit, confirm, and so on. Sorry, but just stop being so negative. This can't be healthy for you.
In taking precautions against thieves who cut open satchels, search bags, and break open boxes, people are sure to cord and fasten them well, and to employ strong bonds and clasps; and in this they are ordinarily said to show their wisdom. When a great thief comes, however, he shoulders the box, lifts up the satchel, carries off the bag, and runs away with them, afraid only that the cords, bonds, and clasps may not be secure; and in this case what was called the wisdom (of the owners) proves to be nothing but a collecting of the things for the great thief.
True or false: LLMs can find security vulnerabilities?
True or false: LLMs can exploit security vulnerablities?
True or false: Any nation with significant resources can make an LLM?
For me these are all true. The last one in specific completely removes the concept of theft and changes it to "enemy nation". Enemy nations can be great thieves, and it's irrelevant, if you don't protect yourself against them you die and everything of yours is stolen.
LLMs are not going away, well, except for you and me maybe. Every state wants to wield digital and informational weapons like this.
Who cannot type "php --version" but can use skills?