It seems like OpenAI are trying to frame this as if AI did all the work here, but it's not really true given what Calif say (https://calif.io/php2shell/):
> 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.
Pure marketing, it's getting increasingly hard to separate actual technical feats of agents from marketing pretentious declarations. Such a terrible environment to navigate right now.
Damned if you do, damned if you don't. Do you prefer they not engage in helping finding vulns at all? Do you not think there are parts of the company interested in security and doing it for the sake of that?
I feel like taking every action these companies take as part of some larger strategy is kind of conspiracy-minded.
The conspiracy theory is suggesting this is part of a larger strategy. I'm just saying it's a big enough company with a wide enough range of beliefs that I find it pretty likely that the direct motives for this weren't to enable regulatory capture.
But they could be, that's why it's a conspiracy theory.
> Calif identified a critical flaw using GPT‑5.6 Sol as part of our Patch the Planet (opens in a new window) initiative
> 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.
keep in mind, the point of most views is a toddler who lies about stealing a cooking isn't corrupt, because they're a child and can't do anything.
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.
They literally open by saying Calif used GPT 5.6 Sol, which is a widely available model. Yes, the person driving did more than "find all the bugs", but nowhere does OpenAI imply otherwise.
> It highlights how AI can uncover security flaws with internet-scale impact
> 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".
In 2024 I saw a reddit post answering the question: AI is going to program insert name App.
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.
I can't think of how it'll really eliminate prompting, but needing to carefully phrase your language already seems in the way out? Wasn't one of the prompts leading to solving fields problems just "look for fields problems and try very hard" or the equivalent
> OpenAI are trying to frame this as if AI did all the work here
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.
Hey world, we created this problem, and it will hurt all of you whether you use it or not. The good thing is there is a solution: if you pay us a lot of money we can try to fix it. The only problem is you’d better hurry because we are releasing the dangerous stuff as fast as we can so you’d better pay us to put the safety measures in place because we are the only people that can fix it and time is running out because, like I said, we are moving fast.
> How can we help identify and fix vulnerabilities in downstream packages while upstream fixes are pending, or backport fixes to older versions that no longer receive security updates?
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?
The commit $1 billion in tokens to "defenders". $1 billion directly to QA for open source projects [1] which would write 100% code coverage, run ASAN and fuzzers would be more effective.
But of course there would be no marketing.
[1] From whom you stole everything in the first place.
I'm as frustrated as the next open source maintainer, but "stole" goes too far. We gave that code away with pride, knowing the world would be a better place.
Few works are released directly to the public domain. It is my understanding that the public domain doesn't even exist in all legal jurisdictions.
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.
You can read code, rewrite the functionality yourself, and sell it, and not be at odds with copyright. Open source does NOT protect against reimplementation. See Java lawsuits.
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.
Speak for yourself. I don't see every model being released under GPL after having read the entire source code as a step of creating their binary artifacts (model files).
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.
It’s marketing, but in a global good will sense of “look how much OpenAI cares about secure open source.”
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.
They found a vulnerability and reported it; isn't that a good thing? Are you suggesting it would be better to just not look for vulnerabilities, hoping that if nobody knows they exist, they can't be exploited?
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?
If they can find the vulnerability anyone can. This whole "patch the planet" is marketing trying to cover up for the danger AI is unleashing on internet infrastructure.
> 1 August 2026: 2 days after OpenAI invites Calif to join its “Patch the Planet” initiative, 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.
No idea what your comment is even about. The researchers that found and reported this did not sit on this for 2 months.
> 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.
Trying to make any story too simple makes it worthless.
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.
hmmm not sure about this one, the disease was already there it just available to anynone (not only for state sponsored actors) and the vaccine is globally available at a fraction of what it cost to build it.
Automated vulnerability detection existed long before the era of artificial intelligence; many people have been doing this for a long time. All the existing tools and documentation paved the way for what AI is capable of today. Furthermore, undisclosed vulnerabilities still do exist—often discovered by state-sponsored actors who exploit them for opaque purposes (frequently operating outside the legal framework of the country in question).
> 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.
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.
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.
> 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.
> 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".
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?
Knowing uv means you can scaffold any python project. docker lets you build any infrastructure kit.
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.
Who cannot type "php --version" but can use skills?
So how many tokens can I put you down for.
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?
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.
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
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.
Just give us the damn code to verify things are working since so many have to use stopgap solutions and want to verify them
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?
It's very difficult to answer those two questions, through all the marketing and fluff text.
Edit: Apparently both found independently
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.
You're not required to disclose bugs, nor is developing exploits a criminal offense (assuming no intent/conspiracy to hack stuff).
OpenAI consistently doesn't know what they're talking about.
How much time will be left in the Defender’s Window between now and an organization getting Daybreak Access?
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.
http://nothingistic.org/library/chuangtzu/chuang24.html
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.