PHP2Shell puts emergency patching under the spotlight. And when you ask an AI assistant to forget a secret, is deleting the original message enough?
Sources & transcript: Episode #51 page
Read transcript: https://share.transistor.fm/s/9ff35c8e/transcript
Apple Podcasts web strips note links: use its existing Episode Webpage action for clickable sources and the transcript.
Automated transcript; may contain errors.
PHP2Shell puts emergency patching under the spotlight. And when you ask an AI assistant to forget a secret, is deleting the original message enough?
Shelby Pettit and Bryce Kunz discuss the newly disclosed PHP vulnerabilities, AI-assisted vulnerability discovery, and the challenges of coordinating fixes before disclosure details spread. Shelby shares the patching guidance; Bryce argues that finding bugs is only the beginning: teams also need to fix their code and deploy it quickly, with a practical emergency-testing strategy.
Then they unpack AfterDeleteBench, research into information that can survive deletion in an AI agent's external memory. The conversation explores summaries, logs, reconstructed secrets, and why a reassuring answer is not proof that every copy is gone. The paper's roughly two-thirds recovery result concerns a specific controlled Qwen2.5-7B-Instruct setting—not all AI assistants or consumer chat products. Zero data retention, private chats, and guaranteed deletion are different things.
Plus: Shelby's Japanese-inspired nail art and Ramune theme, Bryce's Disneyland trip, and a candid comparison of boutique appointments on land versus at sea.
What is your team's emergency patching bottleneck: finding the issue, fixing the code, or getting the fix deployed? Let us know in the comments, and subscribe for more AI and cybersecurity conversations.
RESOURCES
PHP version checker / Codex skill: https://github.com/openai/php2shell-checker
PHP Foundation release guidance: https://thephp.foundation/blog/2026/10/21/todays-php-security-release-and-how-were-strengthening-php-security/
OpenAI: Patch the Planet coordination: https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities/
AfterDeleteBench paper: https://arxiv.org/abs/2610.31366v1
Show: https://aicyber.land
KEY MOMENTS
00:00 Welcome, superintelligence & the subscribe pitch
00:56 PHP2Shell: the vulnerability chain
02:20 Emergency patching & checking PHP versions
04:02 Disclosure leaks & coordinated response
06:43 Preparing for more AI-discovered vulnerabilities
08:05 Find, fix, deploy: where defenders should focus
10:05 Emergency testing & patching trade-offs
12:01 AfterDeleteBench: can AI forget a secret?
13:13 Secrets in summaries, logs & memory
14:42 Zero data retention & private-chat controls
16:03 What would genuine deletion require?
17:10 Don't share unnecessary secrets
18:01 Shelby's nail art & Ramune
18:47 Bryce's Disneyland trip & character dining
20:03 Boutique appointments: Disneyland versus a cruise
21:08 Wrap-up & a final patch reminder
Topics: #AI #Cybersecurity #PHP #AISecurity
Join industry experts and thought leaders as we dive deep into how artificial intelligence is transforming cybersecurity, shaping defense strategies, and creating new opportunities in the digital landscape.
Automated transcript; may contain errors. Formatting/encoding cleanup only; not a full human audio or factual review. Corrections: info@aicyber.land.
Hey, welcome back to the AI cyber.land podcast where artificial intelligence, some people call it super intelligence, but we're not sure if those people have super intelligence themselves or not. That's questionable. Uh, wait, collide with cyber security. Something something. Shelby, the world's greatest co-host is here. That much I know. Hello and dog Bryce. So, uh, we're your eyes and ears for all things AI and cyber security related, the breaches, the breakthroughs, and more.
Uh, so, uh, kick back and let's do this. Uh, if you haven't already, please subscribe. Really appreciate everyone who subscribes. It's a huge help. Uh, so that I can feel good about my life choices. And, you know, you're just generally going to make my life great if you click that subscribe button. >> All right, Shelby. What have you seen this week that you want to talk about?
>> Well, we have to talk about PHP to shell. This has been all the buzz lately in security. Um, basically PHP to shell is two vulnerabilities put together which can allow attackers um to take control of vulnerable servers um web servers. So PHP is used by up to an estimated 70% of the world's websites. Um so the attack surface is massive to put it lightly. Um and the it was first discovered by um an organization called Califf um back in on August 11th.
So why are we just getting the patch now? I'll tell you why. So okay, so it was discovered with uh OpenAI's GPT 5.6 Saul. Um and then interestingly enough in the weeks after so they started working you know they did responsible disclosure started working on a patch for it trying to test that patch with adversarial AI trying to attack it make sure that there there weren't workarounds so that their patch was actually working you know all the all the things that go into that so um interestingly so califf discovered it and it affects like the last 13 years worth of PHP so huge exposure um whiz and open AAI also discovered the this these vulnerabilities independently in the following weeks.
So, they also found it before and these were found with like Frontier Models. So, um I just want to every time I choose a story for our podcast, I try to think how to make it most useful for our listeners. So, if you're listening, we're just going to throw this out there right now. You need to go do an emergency patch of your PHP. Um and I will post a link to the GitHub where there is um what is it called? a codec skill which will do a PHP version checker and let you know which of your uh software needs to be updated so you know where to go apply the fixes.
Anyway, so how to get that out of the way? Need to say the the solution for everyone right now that's listening is go patch these Pronto. Um so an emergency patch was released just on Wednesday. So this is brand new news. Um it was actually so I follow one of the guys who was one of the um the researchers who worked on this patch from the beginning um in with in this like collaboration of OpenAI.
They eventually actually brought in um so obviously the PHP maintainers um then they eventually started bringing in people like Whiz and Cloudflares to figure out how to how to fix this as fast as we can because we're talking about an unau unauthenticated remote code execution. So, you know, they're just trying to reduce that blast radius, trying to stop as much of it as they can because as soon as, you know, it's announced that there's this vulnerability, it's going to be start, you know, people are going to start taking advantage of it.
Um, OpenAI said that as far as they know, um, they think they're the first ones to or like they think that, you know, the good guys found it before the attackers. We don't know of any evidence of it happening in the wild. So, that's really cool. And it kind of shows a little bit of that gap between the frontier models and what's publicly available potentially. But I think it is kind of a narrow gap and it leads to interesting discussions of how much longer can we stay ahead of the attackers?
Are we ahead of the attackers, right? How do you give the right tools to your defenders because it's also going to fall into everyone's hands, right? So, it's a very interesting balance. Um, anyway, sorry, going back. I I follow one of the guys who was working on this project and reading his inside thoughts about it and he actually mentioned that the patch was supposed to happen on Thursday, yesterday, but it had to be moved up a day because people started leaking rumors on Tuesday.
So, they had to bump it up a day and just release the patch. Um, which is interesting cuz they they knew from the beginning that that was a possibility. There are a lot of people working on this. And as we've talked about in previous episodes of our podcast, just the rumor of a bug is enough for these models to be able to find and exploit the bug, right?
Like you just have to have an idea of where to point it at and then throw some tokens at it and it's going to get you an exploit in in some time, right? So, um, responsible disclosure is getting a little bit more tricky and, um, some interesting insight from OpenAI. Their response to this came 2 days later where they said they had a blog post. They're like, "Hey, we're going to have this thing called patch the planet and it's going to be a global coordinated effort of disclosures of critical vans because we are seeing this problem where as soon as an an a vulnerability is known, it can start being exploited, right?
Um, and so it talks about that balance of like having to get all the right people in the room to fix the issue um without bringing in too many people cuz loose lips sink ships and um you know even with the small curated group that they have of this coordinated effort with the the PHP issue they still someone was still leaking like so like just a reminder to all my security researchers out there please be very careful with until there's a patch available you it's just uh just hold your peace you know it might not be very helpful if you tell anyone hints about what you found.
Um and and so yeah, I thought it was interesting. I did read one article that was kind of critical of OpenAI like, oh yeah, never waste a good crisis, right? It does seem a little opportunistic of them to be like, and now we've got this global approach to things. Um while on the other hand, they're also making the tools that help find all of them. It's very complicated, right?
Anyway, um, in the article I read, it calls it the most impactful security issues in recent memory. And OpenAI also said, "This is not the only high impact one they're working on." They're like, "Yeah, this is a big one, but we've also got a bunch of similar findings, and we should expect that the number of these kinds of findings is going to sharply increase in coming months, which hopefully hack patch the planet is going to be like or things like that will be a a good answer to that.
But I think it means people have to revisit their patching cycle. You can't wait till, you know, every every other week or something like that. These are these are emergency patches that are coming out because just everything's accelerating. So that's that. There's PHP to shell for you. What do you think, Bryce? What have you been hearing about it? >> Yeah. I mean, yeah, I mean I think there's a lot of vulnerabilities coming down the pipe.
So, uh, transparently, yeah, I I think we're going to be kind of in a, uh, yeah, patch cycle for probably the next quarter. The um the thing that I think is really important that I think people now are starting to realize which probably should have been apparent from the get- go is that everyone like let's reel it back like 9 months everyone's really focused on project glass wing which is more like hey let's go find all the vulnerabilities that are out there but really where I think the industry should have been more focused and where I think companies and this is what I advise to the people that I talk to is like there's finding vulnerabilities Right.
Whatever. I mean, one could make the argument this PHP phone's been around 20 years. Somebody probably found this sucker in the last 20 years and has been using it. I mean, I don't have any proof of that >> possible, >> but I'm just saying there is an entire market for ODAS where people buy and sell them, right? So, I, you know, it's entirely plausible someone could have made a profit off this or several people at this point.
So the um the the thing I want to say emphasize is like what really matters is um like time to make changes to your infrastructure uh especially code bases, right? I think we're going to see a lot more like dependencies that have vulnerabilities in them that are rolled up and commonly used. I also think that the so part one is find vulnerabilities. You don't really need to worry about that.
Other people are already doing that for you, right? Unless it's your I mean if it's like a library or a piece of software that or a piece of technology that other people use, you don't really need to worry about that. If it's your own code like in house that nobody else has, then yeah, maybe worry a little bit, but it's still not as big a concern as these public things that are internetwide.
>> Then step two is how fast you can change your infrastructure. And by change, I mean like make the change to the code. And step three in my opinion is like now that you have a known good state, how fast can you deploy that and change all your internet facing or internal facing infrastructure as fast as possible? Cuz it's one thing to like hey we have the fix but in a lot of software deployment models that's not good enough to actually not be vulnerable because you have to like make sure all your deployments are also running that latest version of the patch software.
And so really of those three columns, find vans, fix code, deploy, like the the latter two are really where you should be focusing your efforts because if you can just get really fast at that, and I I I know this is trivializing trivializing it to a certain degree. Um because it's not that easy to get fast at that to be honest. Uh like specifically, the more money you make off your code, the more you're going to want to do testing on your codebase before you do a deployment, right?
There are a lot of companies where that testing of a change of the software takes them more than 12 hours, right? Like could possibly take them more than a day, right? And you just don't have that. You don't have that luxury anymore in this world of mass AI agent exploitation. If you're going to wait a day to go through all your regression testing on your code to make sure it's not going to fix some like you're fixing one thing, so you might break something else, right?
Then you're not going to be able to like you're going to get exploited, you know? So, >> so you say we should do yolo mode for deployments. >> Well, I think you need to like fundamentally think like what is our risk posture? What is like like the barebones set of tests that we must run before we push like an emergency patch out? And yeah, we'll run the whole test suite like let it go to make sure we don't break something.
But, you know, you're going to want a tiering strategy for your test suites. like you're going to want like these are the ones we have to get through and then we can release deploy and then we'll have that tailwind of like another 24 hours of test to make sure we didn't break that >> and I mean it's kind of like a it's kind of like a trade-off like I heard a bunch of people with the PHP uh like they took their websites off the internet completely like they're 100% down while they were patching in fact there's a rumor which I don't have evidence of but there's a rumor that pretty much all of North Korea's systems went down so they probably just pushed out and mandate like, "Hey, take them offline till we know everything's secure and then bring them back up."
So, so >> an outage is better than not owning your own servers. That's fair. >> Well, that's what happened in this case. Several companies, like US-based companies, decided, hey, we're just going to take the outage because that's better than, you know, dealing with the instant response bill, right? So, >> yeah. >> If anyone knows of any of those companies or saw any websites that just straight up went down, let us know in the comments because I'm curious now.
>> Yeah. Yeah. Yeah. Yeah, there's definitely a few. So, I comment below if you know any names. Uh yeah, I mean it it's just this stuff's always tricky. Um it's hard to keep secrets, right? So, when they're doing these coordinated patches, I do like the patch patch the planet. That's pretty good marketing. Whoever came up with that, give them a bonus. So, I like that. >> It's got a nice ring to it.
Except I keep wanting to say hack the planet, which I know is a totally different thing, but we'll get there. >> We're doing both. We're doing both. Uh what you see in the news, Bryce, >> in other great news this week, uh new research came out and there's a new benchmark called the after delete bench. And this benchmark essentially what it's going to do, well, let me let me back up a little bit so I don't jump the gun too much.
Okay, so imagine you like work at a bakery and you make awesome cupcakes. You're living the dream. And then you're using an AI assistant. Let's say you're using like an Alexa or like a Siri or something like that, you know, and it's listening to you as you're operating uh and you tell your coworker the code for the alarm system, right? So that they can alarm the system.
Now you realize when you come to work the next day, it's probably not a great idea that the AI assistant knows the PIN to the alarm system. Like you want to be in charge of giving that to people. Don't let like the AI agent do that. Um, so you tell the AI agent, hey, forget I ever told you the alarm code. This after delete bench essentially tries to see what is the percentage of time where the AI models or the systems and harnesses around the models when do when do they actually delete the secret?
Right. >> Oh, and >> yeah. And so uh the results are uh they're not great. So it turns out twothirds of the time, so two out of three times, if you ask the AI assistant like the next day, it will still be able to tell you back the password or secret. So >> Oh, >> essentially what happens is like anytime your chat goes idle, these harnesses or operating systems that sit around the models essentially are doing summaries of your chat sessions.
And so those secrets are getting saved to the summaries. And then there's also like chat logs that could be used for like debugging purposes and other places where it could potentially store the secret as part of just its normal operating procedures. And so when the AI agent is asked subsequently, let's just say like a new employee shows up and ask the AI assistant, hey, what's the PIN for the security code, it will because it's trying to get you the PIN, like that's the goal, it'll go through every known data source it has access to in order to try to reconstruct the PIN to you.
And they even went as far as to break up secrets into chunks, right? So they were like the first number is five and then like later on in another chat they'd be like the second number is six and then later on and then they would be like all right they'll delete the secret and the AI assistant still like like would gladly recreate it. Right. So >> really.
>> Yeah. So >> so I mean it doesn't happen every single time, right? But, you know, two-thirds out of three, that's most of the time. And the the biggest thing here is like, which I think is kind of a bigger issue than just what I'm talking about specifically here, but it's kind of this ZDR concept, right? Which is like does the system you're operating have a zero data retention policy or ZDR policy.
Um, so like if you go to like open router is a good example. Open router routes coms to different AI models. And if you say in their settings like, "Hey, I only want to use zero uh basically zero attention uh models," it will only route you to models that support that type of privacy uh features. Now, if you're just like the an average user and you're using chatbt.com, uh some tips for you is uh most like chatbt, Gemini, Claude, they'll have like a little icon in the upper right hand corner of their web interface that allows you to open a like incognito or private chat.
Um so that will be a little bit more secure than just doing a normal chat. I also saw uh that chatgbt is releasing a feature where you could actually set an additional pin on a chat uh which means like if someone sits down and your chatgbt.com is open and when they go to click on a chat history that specific chat history could have a pin that they additionally need to input to get access to it.
So there are there are some like compensating controls that are out there to help you. Um, but I mean really what we need is we need kind of these operating systems or carnisses that sit around these models to be a little bit smarter. Like when you tell them like hey I want you to forget this specific item uh you know in the future hopefully they'll get in order to get higher marks on this benchmark they'll have to also go and delete the other artifacts that are associated with it or at least redact those so that you know when you go to do a delete it actually deletes the secret.
Um there were some pretty interesting results in this white paper too uh like models stating back that they didn't know the secret but then when you pressured them they 100% knew the secret right and things like that. So so it was generally the kind of TLDDR was like the smarter the model is the more likely it was to actually destroy the evidence of the secret. the kind of weaker the model was, the more likely it was to hide the secret somewhere or not realize that it had still had access to the secret.
>> And uh yeah, so those memory features that make these systems get to know you and customize and tailor the responses to you also are the same features that are going to you know retain these secrets uh which potentially you know is a security risk. So anyways, it's called the after delete bench. Uh, we'll put a link to the white paper below if anybody wants to check out the research and, you know, when in doubt, maybe just don't tell the AI their secret, you know.
So, there's my there's my pro advice, pro tip for you. So, >> the best way to not leak your secret is to not tell it. >> Yeah. Yeah. And if you have >> way to not get your web server hacked is to not have it online. >> Yeah. Just pull that PHP website off the internet. You didn't need that anyways. who really the scary part is if you're running a PHP web app, odds are high that the person that developed that is no longer at your company.
So, uh so good good luck getting that fixed. So, uh >> out there everybody. >> Good luck. Yeah. Okay. All right. Um uh have you done or seen anything fun in the last week, Shelby? >> Yes, I have. I've been so excited to show you. Now, I know that the overlap between people who are interested in cyber security and AI is probably a small overlap, like if we do it on a ven diagram of people who are interested in acrylic nails, but here they are.
>> Oh, man. >> Beautiful nails. And they're like Japanese kawaii style. >> This one is like a specific brand. Oh, it's upside down. A specific brand of Japanese drink. It's probably not going to focus on my thumb. Oh well. This one I'm going to drink. The one with the little marble that you have to push down to open it. >> Oh, yeah. I'm can't get a job.
>> That's pretty crazy. Cool. >> But really fun. Anyway, yes. So, there are some of us who like both security and nails. I'm so happy. What about you, Bryce? What have you >> pretty sweet? My Yeah, my daughters love the fancy nails. They're crazy about it. Um, yeah. Uh, new in the world of Bryce. I went down to Disneyland. I saw Mouse and other characters. It was super magical.
I think uh Disneyland's kind of like like just to be honest, it's kind of like a polarizing experience because I could very easily see how you could take the family to Disneyland and have a horrible day. Uh just cuz like I mean it's just so brutal if you don't know how to operate inside the theme park. Uh especially with like you know some other people that do they're attending the park at the same day.
So, I don't I don't like universally recommend people go to theme parks in general, but cuz like honestly, it' probably be way less stressful to just like go to the beach and give your kid a a plastic cup and a plastic spoon and let them go crazy with the sand, you know? And I feel like that equivalent would probably have way less stress and way more joy in your life.
But if you are going to do the Disney thing, uh we did the Minnie Mouse character meet and greet at the Plaza Inn at Disneyland. That that was pretty good. Uh I thought that was like good value. It was about $40 per head and they gave you like brunch with it and all that. So, >> okay. >> Yeah. And things that I would not recommend doing, and I know this will probably go in the I'll get some hate for this in the comments below cuz there are a lot of Disneyland people out there in the comments below.
Uh, dude, the Bibbid Bobby Boutique at Disneyland. So bad. Like seriously, the worst. >> I know. Why do people love that? I don't understand it. It's like I pay them a lot of money and then they make me wait a line and then they take me to the next line. I wait in and the next line I wait in. It's like you have an appointment time, right?
Like that's a definitely a pet peeve of mine. If I have an appointment time, I expect like when the appointment time is the appointment time that we get rolling at the appointment time. >> That's a good point. >> Appointment time. >> But I feel like that's the theme of Disneyland. It's like magic and waiting. >> But I I have to say like I like I've done the Bibby Body Boutique for my daughters at the on like a Disney cruise >> and >> like that's a 10 out of 10 experience. like, >> okay, okay, >> you show up at the appointment time, maybe you wait a couple minutes for them to get ready and then they do the whole thing all together and then you walk out there and you're done, right?
It's not like go to >> spot A, do something, wait in another line, wait in another line, wait in another line. I don't I did not I I thought the Disneyland one was a ripoff. That's my personal opinion. I don't think you're going to change it. >> We've been warned. >> But my kids loved it, so whatever. Uh, and that's a wrap on today's briefing. If you like wasting money on things that you shouldn't spend money on, you should click the like button.
>> If you think >> Yeah. >> If you think Bryce is an idiot for spending his money certain ways, click that like button. That's the way that that will be the way I'll know. I need an intervention. And uh All right. Try to go past your servers. Thanks everyone. See you next time. Bye >> bye. >> You have been listening to the AI cyber. Land podcast. Join the resistance at AI cyber.land.