Risky Business #856 -- 8-bit encryption, PHP megabugs and more
00:00Hey everyone and welcome to risky business. My name's Patrick Gray. We've got a great show for you this week. There's plenty of news to get through and joining me to discuss it this week is my colleague James Wilson. And in the third chair this week we're joined by Mark Coory, otherwise known as Salt Ice, on Twitter and by day he is a security engineer. I guess by night he's still a security engineer, but also does some YouTube stuff and he's.just generally pretty sensible when it comes to all things cyber security. So we asked him tojoin us to talk through the news this week. Mark, thank you very much for joining us.Thank you for having me, a huge fan of both of you, Patrick and James. Happy to be here.Longtime listener, first time caller, right? So that's great. And this week's show is brought to youby Trail of Bits. And I didn't do this week's sponsor interview. That was James who did it.He spoke with Anish Naig, who talked about some of Trail of Bits' work inwith Open AI, the Patch the Planet stuff, and they've actually created a couple of
01:10plugins or sort of skills files that people can use to find bugs, right?So they're going to be chatting about that later on in this week's a sponsor interview,but first up in this week's news, yeah, so this is ominous.There is apparently a pair of CVSS like 9 plus bugs coming in P.H.Core.And this is what I'm loving about the AI Age, right?Is we stopped seeing certain types of bugs.
01:41We stopped seeing bugs in WordPress core and, you know, everything moved to the plug-ins.But, you know, those sort of bugs are back.We stopped seeing bugs in things like PHP core.And now they're back.We saw a bug in Drupal recently as well, right?You know, but PHP core, like when you think about how much of the internet runs on a lamp stack,this is like, this could be like a meteor hits Earth moment.apparently details of the bug and the patch are coming out like later today.
02:08So it could be a huge impact or as often happens it could just slide right by with minimal impact, right?And who can say what will happen?But what's your feeling, James, on how this is going to shake out?Because I figure it's like a coin toss between nothing noticeable or internet on fire, which is fun.Yeah. I kind of think this might be internet on fire just when you think about beyond PHP,he, you know, WordPress, Mugento, Larravel, all the things, like you said, the lampstack's so oldand so prevalent. Yeah, it could be, could be an interesting week coming up. Yeah, and we saw a,you know, really bad bug that was patched over the weekend in like, you know, a kernel-based
02:47virtual machine software, right? Like that, that was huge, didn't seem to amount to much, but I guessthere's like a handful of really big providers who needed to sort that one out. Mark, you havenoted here that there's a lot of tweets telling people to patch now, which is funny because thepatch isn't actually out yet. Yeah, yeah, on Twitter, everyone's like, you have to patch right now.At least I saw a few tweets like that. It's like, well, maybe wait till tomorrow until the actualpatch is out. Seems the important part of it. There's also been some funny tweets around,oh, PHP, nobody uses PHP. Why are you using PHP? You deserve it. It's like, oh, actuallya large double-digit percentage on the high end of the internet is running PHP. So, likely,
03:27likely a lot of people affected by this.Yeah, I mean,PHP makes Java seem obscure, right?And I don't think everyone necessarily knows that.So it could be a fun day tomorrow anyway.I feel like at the moment,working in security is like being someone with a telescopewho's seeing these huge meteors just like slide by
03:48just grazing the atmosphere, you know what I mean?But not causing any damage.Like one of them is going to hit us eventually.So let's see if someone,one uses like GLM 5.3 to worm that one tomorrow in which case oh that's going to be a fun day.Now look speaking of some other stuff, this is just funny. That's the only reason this one's in here.Some certificate authority apparently says a bunch of their intermediate certificates turned out to be generated by keys that wereI guess you would say they didn't really contain much entropy given that they were one of 256 possible values.
04:26Mark, you've been tweeting about this today.Please tell us about these 8-bit private keysbecause it's somewhat hilarious.I was fascinated by this onebecause initially it didn't make sense at all.So the private key associated with these intermediate CAs,C-A certs, was one of them was four,the number four,
04:48and you can even see it, like it's just like 60 zerosand then a four at the end.And it turns out that what likely happenedwas that the random,number generator was the wrong one maybe because instead of having around 360 or 384 bits ofentropy there was just eight and so all eight of those certs had private keys that were one of themwas four then you had 56 the largest one was 237 all of them less than 255 I mean it does it doestend to suggest something has gone wrong in a in a tool in a chain of tools somewhere that
05:26that is, you know, generating these things.Yeah, and why?Like, how could that, you're a CA?It seems like you would have to have a good understanding of how to avoid this to become a CA.Yeah.Well, I mean, you would think that, but like the length, like the sort of stuff we seeout of CIA is on the regular, you know, apparently, apparently not.But yeah, this is some company called Emudra, I guess, is how you would pronounce that.
05:52And, yeah, so I think Mozilla's already revoked those search.I think six of them were for TLS, two of them were for code signing.James, any thoughts or anything interesting here that you wanted to add to it?I'm a little bit curious as to how the researcher found that these private keys had such low entropy,because the keys weren't leaked, they weren't stolen, but he somehow discovered and reproduced the results here.Well, I think we're having a joke about, you know, calling this brute forcing when the key space is 256 bits.It's less brute force and more like, you know, tickling or something.It's a mild tickle.
06:24I mean, was that it?Was he just running like something?on test of like what happens if I produce a certain and suddenly it happened to be valid forthese ones? I don't know. It's an odd one to like, how did you stumble upon this? Can you imagine ifyou're doing a sequential brute force and you get like, you know, eight hits on the first.Exactly. You get eight hits on the first 256 attempts, you know, that would be funny. All right. Somoving on, let's have a quick chat about this open code breach because this was kind of interesting.It looks like whoever was behind this, what they were doing is compromising open code console accounts.
07:01Now these weren't, you know, it wasn't CLI based accounts, so it was people who were using open code console.And then what they did is they used those accounts, those accounts accounts with the Frontier Labs to run a bunch of queries to do distillation.So basically someone breached open code console in order to steal token.and to conduct some sort of distillation attack against frontier models, which, you know, I mean, I see this and I sort of think, you know, nice. Let's get you on this one first, James.Yeah, look at this and just go, this is the new norm, right? These tokens are so valuable and so useful. Of course, if you're in the position of holding them, you're going to be, you know, target number one. That said, there's some details here that don't quite stack up just yet. Like I haven't seen and I went looking to try to find evidence.of how they're actually determining this was a distillation attack.I think it's a reasonable assumption,but I haven't seen logs or transcripts that show exactly what they were doing with this.
08:05The other thing that I think is a little bit odd is some of the commentary around,you know, we were getting hammered by, you know, 48 hours solid.They kept attacking us over and over again, and we had to deploy our own defensive agents.You know, it sort of reminds me of Clem from Hugging Face after their incidentand someone said, hey, you know, why didn't you have AI running, you know, AI against AI?And he had a very calm and good reasoned response where he said, look, that's just not the how that happens, right?We don't have AI playing sword fights in attack scenarios.But anyway, look, it's interesting to see.
08:37And hopefully they do publish a really detailed post-mortem so we can get some more details around this.I mean, they did say, though, that in this case as well, they had problems trying to use the frontier models to do instant response.so they had to switch to the Chinese models.I mean, this just keeps happening over and over and over againbecause the guard rails are kind of nuts.It's funny just when you were talking there about like,well, was this a distillation attack?It reminds me, you know, there's that meme people send around,
09:04which says, like, everything I don't like is woke.And there's a guy pointing to an engine in a carand he's like, this engine is woke.It sort of feels a little bit like that at the momentwith the American Frontier Labs, like everything I don't like is distillation.It is a little bit like that.Mark, you had some thoughts here,You also noticed that open code, they recently patched a pretty serious, serious bug, right?
09:27Yeah.So as I was looking into it, trying to find any more information on it, I found that there was an article published today, October 20th, on how open code the client, the harness on Windows and WSL had an RCE where if you connect to malicious MCP, you're just vulnerable to RCE.and apparently OpenCode or Anomily Team patched it a couple weeks ago silentlywithout really disclosing anything or issuing a CVE.And that just makes me think about their maturity in terms of transparencyand how they respond to incidents.Because as James mentioned earlier, comparing it to the Hugging Face incident,they were very transparent.
10:10They had an excellent post-mortem report.I hope to see one for this incident,but we haven't really heard much or seen any.yet.Yeah, I feel like Hugging Face were really transparent because it was an awesome way to shake downopen AI for tokens though.But yeah, no, I take your point.I mean that this vulnerability that you had, we're just talking about, I mean, it probably
10:33doesn't connect to this incident, but the fact that they were like, oh, nothing to see here,CVE, what's that?You know, probably not a great sign of how they're operating.What else we got here?DHS in the United States is banning deep sea.from federal civilian agencies.I don't think this is surprising at all.The way the show works for those who are not familiar was,
10:55you know, we develop a run sheet through the weekand we all put comments on it as we, you know,get closer to the show day.I think there's a comment here from you, Mark, saying,well, you know, why are they not self-hosting these things?And my guess is, you know, for the average Fed in the US government,like trying to get budget approval for inference hardware, like it just ain't happening, right?So I'm guessing they are banning
11:17people from actually using deep seek accounts, which I think is, you know, I don't thinkU.S. federal government agencies should be using Chinese cloud AI. I think that is a, that is not a good,that is not a good idea. James, would you agree? I would certainly agree with that. I will say, though,that when you actually read the notice, it's more than just banning the internal use of deep seek.This is deep seek, deep seek affiliated products and third-party products and servicesdeveloped using deep seek implementations. Well, how are you going to enforce that?I don't know. I mean, but it's just a, it's a, on one hand, it's an oddly specific notice,because why just deep seek? Why not GLM and Quinn as well? But then oddly quite broad.
11:57It's like, you know, I won't buy anything off you even if you've used deep seek,not that I'll ever know that you did. It's...Well, not that even, and I think there's a point Mark's drive out in the notes here,it's like, you know, if you're using some sort of LLM proxy or router,like you might not even know that you're using deep seek, right?Yeah, don't worry. We're not using deep seek. We use open router and open code.That's what we used.Exactly, right?
12:19So, Mark, what are your thoughts here?Yeah, my initial thought was, like, are you not self-hosting?So if you are self-hosting, why would you care?I mean, there's a separate conversation about, like, the model having bias.Maybe that's why they want to avoid it.Because the model is a communist.Right, right.So if you're self-hosting, it should be fine, right?
12:38And if you're not self-hosting, why are you beaming all of your government docks to servers abroad?That's also weird, right?I mean, if you work for the part of the U.S. government that is responsible for cataloging the history of the Tiananmen Square incident, the Tiananmen Square massacre, you know, like maybe don't use Deepseek on a self-hosted thing.But it's hard to imagine outside of that.Why, yeah, I mean, considering it is used, Chinese models are used to do exploit development.Where the exploits, I wonder.So that's an interesting thought.I wonder if there are exploits currently in use by the U.S. government that they bought from contractors who may have even.
13:17use deep seek to do some of the exploit development that'd be interesting i think almost certainly by thispoint yeah yeah that's it uh now look we had a really great uh statement tweet from uh rob joyceuh you know friend of the show uh connecting to something we've been talking about on the shownow last week i i said that you know coming back into the office after a couple of weeks offi had seen just heaps of breaches more than usual feels like things are really ticking upyou know brad arkin uh his reply to that was well you know we may have seen a doubling an incidentbut we haven't seen an increase in incidents that sort of commence here with the increase of like, you know,public vulnerability reports and whatnot, which is a completely reasonable thing to say.
13:57But Rob here seems to lean towards thinking the breach apocalypse is here as well,like things are really ramping up at the moment.He links through to a Reuters report about some Chinese-aligned hacker or, you know,Chinese-speaking hacker who used AI agents to own like, you know, heaps of websites automatically.but he also said something interesting,which is that IR firms are saturatedand turning away work.Now, there's nothing mentioned about that
14:24in the Reuters article.And a lot of people were replying to Robsaying, well, there's nothing in the article about that,you know, citation needed sort of thing.You know, I know Rob reasonably well,and he's not a shoot from the hip kind of fella.But I did get in touch with himand we just had a bit of a back and forth,
14:43you know, messaging about this,and I'm like, I'm guessing this is somethingthat people in IRtelling you and he's like yes it absolutely is so you know I generally feel it's pretty safe to takestatements of fact from from Rob Joyce pretty seriously um but I don't know man it does feel likewe're at the precipice of something but we are just gonna have to wait aren't we but James whatis there anything particularly novel or interesting about this you know and what a difference a yearmakes because a year ago we would have been talking about you know a single hacker using
15:13AI to you know I think it's deep seek and clawed in this case wasn't it yeah toown a bunch of websites like that would have been a you know stop press's moment and now it's likeit's uh you know it's just a regular wednesday i look i think this all comes down to a visibilityissue right we we hear about the volumpox eclipse and the bugpocalypse because there's so muchglory to be gained from getting the cves and getting on those advisories there's very littleglory in ending up in a breach notice and so i imagine there is a huge amount of breaches occurringthat are still either coursing through i are going through internal sort of triage thatnot yet, you know, publicly known outside of just firms that have to do a certain amount of
15:51publicity within a shortened amount of time. So, yeah, look, I have no doubt that there iswreckage, left, right and centre. We're just not seeing it yet. But those, those IR firms turningaway work would be a very good leading indicator. Well, that's exactly what I was going to say,like, you know, running out of IR capacity is a little bit like, you know, the yikes emoji,a little bit like that. Mark, you are the one of us, you know, here, you're the only personhere who's actually a practitioner who's out there at the coal face, you know, youI'm guessing to a bunch of people, you know, in New York, New York where you're based about what'sgoing on out there. What's your feeling for the state of things at the moment? Because it really
16:28feels like, you know, because I monitor the news very closely, obviously, really feels likeover the last month or something, there's been a pretty serious uptick. There's definitelybeen a serious uptick. And I completely agree with James' point around how people like to talkabout breaches and increased offensive capabilities. And we're seeing a lot of thediscussion around that. And on the other hand, there has been also a very real and significantimprovement on the defensive side, on the blue team side, both for fixing your stuff,detecting vulnerabilities, but also for incident response, for knowing how to respond to anything,there's good improvement there. But what's sad is that given all of that improvement on the
17:14offensive side and given that we're seeing more attacks,So far, it hasn't really, anecdotally, maybe, and from the people I talked to, it hasn't correlated with increased security budgets.Even though the dollar spent today, you can do so much more with it, even if it's just like tokens, but like with one extra security engineer headcount on your team today, you can do so much more.But I don't see teams growing in terms of security engineers or detection engineers on the team.In fact, some of them are shrinking, thinking like, oh, I can just get clawing.to do anything, but you still need someone to drive Claude.Here's some tokens, peasant.Here's some tokens, nerd.
17:55Now, stop bothering me, is the attitude.I mean, let's see if that holds.I have a feeling, though, that if we're running into capacity constraints withincident response firms, we're going to start hitting capacity constraints in otherparts of the industry as well.So, I don't know, maybe you want to hire those engineers while you actually can, right,before we start hitting those hard caps.Speaking of engineers, this open AI guy, this is such a weird story, right?
18:22So I think this was just breaking last week as we were recording the show around that time.This guy is a Saudi Arabian guy who was working for Open AI,he put a bug behind a television in a conference room.Now, what makes this interesting is, you know,it's been reported as this big thing, espionage,but he's been like asking Open AI's models like,where do I put a listening device and how does it work and how to how to how to how to how to do corporate espionage and whateveruh wired has some really good reporting here where they've looked at the court documents and stuff
18:56it turns out he's previously worked for a saudi venture capital firm he's offered consulting servicesto saudi arabian companies and apparently he's friendly with like a saudi royal and there's umfrequent communications via WhatsApp with this member of the um of the saudi royal family uh both lastyear or this year and you know this royal uh allegedly uh was trying to recruit him for some sort ofSaudi arabian uh i initiative and said look Saudi needs you uh more than the US and the individual likehit hit hit a 100 react on that one right so you sort of get the impression here that this feels likethis doesn't feel like you know state sponsored espionage per se but the line's a little blurry it'slike, hey, you know, we're spinning up this AI initiative. You've got to come back, do it for your
19:47country sort of thing. And then all of a sudden he's sticking listening devices in conference rooms.So I sort of feel like this is mostly like just typical like employee secrets, theft, but with like a littlebit of a shade of like state aligned espionage on it. Like James, did you get that vibe as well?Like it's pretty clear what's happened here and what was motivating the guy, I think at least.Yeah, I agree.Like this just comes across as a very, very foolish attempt to gather some data that's going to be helpful for all his next adventure moving back to Saudi Arabia.Like I don't get any sort of deep state tradecraft.You know, as we were joking in the notes for this, the real tragedy here is that the AI's advice was to use like an ESP 32 board, wrap it up in toilet paper and jam it into the back of the TV.
20:38Like, what?We can do so much better.You can do so much better.And this is what Trump calls super intelligence.Man, it's so amazing.It's so amazing.I'm getting the press releases now where all these large companies are now referring to it as super intelligence.And I'm just like, aren't you embarrassed?
20:57Like, aren't you embarrassed?Yeah.Are you embarrassed?Are you embarrassed?You should be embarrassed.I think I sent a screencap of one of those press releases to a mate, followed by a image of one of those solitary chairs in a hoax.hotel room next to a bed.A little bit what it feels like.
21:17Now, look, speaking of AI safety, though,Anthropic is shopping around its own safety bill to lawmakers.I don't want to get bogged down talking about this one.Politico got their hands on it somehow,and their proposal allows,it's funny, James, because you zeroed in on the same things I did on this one.Their proposal allows the government to block a modelonly if it poses a foreseeable imminent risk of more than a billion
21:40dollars in damages or of killing or seriously injuring more than 50 people. So I like that they areobsessed with model safety, but they will allow their model to kill up to, you know, kill 49 people,that's fine if there is an imminent risk that it will kill 49 people, but not 50. That's where theydraw the line. Yeah, clearly this is like the anthropic version of the trolley dilemma. The switch goesone way. It's 50 people. The switch goes the other way. It's 49. And they've solved the problem.The dilemma's gone now. 49's fine.Yeah. I mean, it just is funny, though, that they talk out of both sides of their mouth, these companies, right?Where they talk about being safety obsessed. They put ridiculous guardrails on everything.
22:18But when it comes to the crunch, what we are seeing at AI at the moment is an epic land grab.You know, this is where, like, technical debt came from in the cloud ecosystem.This is why Azure has so many issues is because Microsoft was trying to get that market share.They don't care about the safety. They're not trying to slow down.It's just go, go, go. We've seen it time and time again.Same things happening in AI.They don't care about safety.They are pretending.
22:42And they like making their models seem scary and super capablebecause it helps them raise money.So let's just be honest about what is going on here.Now, this one, Mark, I want to talk to you about this one.And you've both got some interesting.We've all got our opinions on this one.But people are trying to do sort of SEO-style gamingwith political messages designed to kind of full
23:06chatbots, right? So, you know, you might be a voter in a U.S. election and you ask, you know,who should I vote for in the upcoming election? You know, here's what's important to me. And, you know,we've had relatively pure results coming back to us because this is a relatively new technology,but it's my concern that some of these SEO types are moving in on this now. And that'sapparently what's happening, according to this piece from Cyberscope. But how's your feeling,What's your feeling for where that's going to go, Mark?We'll start with you because you're the American,and this is focused on the American election, upcoming midterm.
23:43Yeah, so this one, first, it's not about training the data.This is about, like, web retrieval.Once you prompt, ask it something, it goes and fetches sources online.And this has been a problem for a while, right?Like even Google, right before the Gemini summaries at the top,for me, Google was unusable because it was all like SEO,optimized slop at the top, like more than a page of just like useless.Slop optimization. Right, right, right. And so it was a problem. Like Google was unusable. You
24:17couldn't find anything of quality because it was all, well, AI generated stuff was just at thetop. And so now this same tactic can be used for your agent to get it to fetch a specific article.And so what are the ramifications of that?I don't know if there's a good solution to that yet.Because as a problem in general, how do you vet the quality of an information source,which is also applicable for even 20 years ago, right?Like how did you know if a news source was credible or not?Yeah, I mean, funnily enough, this is why I'm an AGI skeptic,
24:57is because how is an LLM in a training run supposed to really know what's rightand what's wrong? Well, funnily enough, that was exactly the chat that Josh Devon fromSondera and I had that we published on Monday. They don't know, but more to the point,they don't care what's right or wrong. That's just not something that is, you know,really an emergent property of an LLM. I mean, that's a bit of that could be a problem later.I mean, I guess that's why I found this story an interesting one to talk about, because, like,don't we want them to kind of, you know, be good at that?Well, we certainly do before we give them opposable thumbs. It comes back to that, right?
25:33Exactly. All right. What else have we got here? We got an absolutely incredible story here in front of us from Brian Krebs over at Krebs on security.So we're now looking at a, he's run a story basically saying that the FBI suspect a ransom negotiator who works for a prestigious American firm may have actually been involved in running Shiny Hunter's leak site, which is just amazing.And there's a whole bunch more detail about this Edward Dobrovsky guy.getting arrested and it's like it's almost too much to summarize here i actually did have a chat withbrian the other day after he published this one he was just emerging right because he was locked infor like four or five days you know he sent me the story i read it and then we were having a chat about itand he's just like yeah man i was locked in and i'm like i can see why like this is dense this storyis dense if you want to really get the skinny on what's going on here um yeah read read his
26:28article, but it looks like this new suspect, really the thinking is he was quite, he was quiteinvolved. But I really wonder, right, about all of this. Like, as I said to Brian, I said, look,imagine you're not so bright, right? This is your starting assumption. Imagine you're not sosmart. And you are working for a company that does these sort of negotiations or helps peoplerecover or whatever it is. And by doing a bunch of forum work, you mind up, wind up coming up withsome, you know, establishing some good contacts with a group like this.And you sort of wind up becoming a bit of a go-between,much like these guys in Somalia who act as go-betweens between the pirates and the companies
27:09who own the boats that get seized or their insurance companies or whatever.You know, and after a time, you stop thinking really, you know, you're just in your niche.You're thinking, wow, I'm in this really good niche.No one else is here.I am very smart because I've established myself in this niche.and you're not thinking, gee, am I actually just acting as an extension of this criminal group, right?And so I don't know what's happened here, but that's why I'm keeping an open mind because, as I say, like, it's entirely possible these people just, hey, maybe they're just not so bright.James, what do you think of all of this?
27:45I mean, it's just a crazy story.It's a crazy story.And I think the thing that I found fascinating is how it's evolved my thoughts around this to Broves.character. I mean, I remember when we looked at this last week, we sort of looked at, you know,the profile photos of the guy and just went, he looks so square. There's no chance he's actuallyinvolved in this. They've just, they've happened to snap him up because he traveled from Ontario toPhiladelphia for a conference and he's got some loose connection, whatever else, right?And then maybe it's a good way to collect some intelligence, choke off the money supply,
28:15whatever, right? There's a bunch of reasons, yeah. He knew how to talk to shiny hunters,therefore, let's go talk to him. He's in the country, sure. Now, you know, the, the, the, the,The Krebs read is so great because it's like watching a plot of a movie unfold.And there's this section where it gets to DeBrovsky earlier this year filed a lawsuit in Ontarioaccusing a former business partner of essentially sidelining him and cutting him out of the businessand then being involved in sort of with this third actor in terms of having a much close relation to shiny hunters.And it's just like, whoa, hang on.This is a lot more of a murky involvement with this when you've got sort of these lawsuits being
28:54raised and all this are leased.I mean, if you're trying to go derp mining,you've just struck a rich vein of derpall around on this, you know.Yes. Yeah, yeah.So look, honestly, I don't know what to make of itother than, gosh, this needs to be turned into a movie,TV miniseries on Netflix at the minimum.
29:13Well, and there's some novel techniquesused by the police in all of this.Mark, you know, why don't you tell ushow the Dutch police got around signal encryption?because it's a, you know, it's an interesting, very interesting technical technique.They needed, they got incriminating evidence about Dubrovsky that was on signal,and you can actually easily decrypt signal communications,and the solution is you wait for, when you see someone on their signal,
29:44you can throw a flashbang on them, and then you go get the phone with signal open.Flawless.Yep, that's it.They flashbanged.They flashbanged that guy they arrested.Pepi Inventus-up in the Netherlands.And he was the one, I guess, was chatting with Dubrovsky or whatever.But yeah, they flashbanged him because they saw that he had signal open.
30:04I just can't wait for the next iPhone to be released that's got like this new featurewhere the ambient light center detects a sudden spike and locks the phone for you immediately.Flashbang detection puts it in BFU mode, right?And it means that someone had binoculars just like staring and waiting for the right moment.Or they had a camera in the room, which is just as likely.but you know fun stuff um what else are we got yeah look there's a whole bunch of these umfridges like we spoke last week about this you know the the the attacks against uh free free freezes andfridges in um europe uh it looks like yeah there's just so many of these devices still online but um james
30:46you noticed something here which is that the devices that were being attacked weren't actuallyindividual freezers. It wasn't your typical IOT stuff. It is like a device. It's still an IOTdevice, but it's like a controller that's designed to control all of the freezers. And you think itmeans that it's a little bit less forgivable that this attack was possible that these deviceswere not being managed correctly because you would think, you know, surely a device that's designedto manage all these other devices should be monitored and looked after properly, I guess.Well, the insinuation last week was, you know, look, this is,what happens when when OT devices fall through the cracks, sort of like framing this as these were little
31:25black boxes jammed inside refrigerators with a 5G modem and, you know, maybe a refrigeratorcontract, a contractor deploys these so that they can do proactive maintenance. And like,if it was that sort of out of sight, out of mind hidden device that corporate IT might not have evenbeen aware was there, then sure, I can buy that narrative that these things do fall through the cracks.This device that was attacked, yes, it's an embedded device insofar as it's not like an app on Windows ormacOS but this it's it's a control panel with a screen that is is essentially orchestrating andcontrolling and managing all of your fridges and freeges throughout a supermarket and it's the kind ofthing that's going to be sitting in the manager's office and it's going to be visible and it's
32:04going to be part of the store running procedures and i actually i actually looked at the at the websitefor this thing and it is something that's designed to be embedded in a wall yeah right so it is stilllike so that's i understand what you're saying but it's still got the feeling of something thatgets installed by a refrigeration tech and forgotten about, I guess.But the moment, that's exactly the crux of it.I don't think this does get forgotten about because this is what the store managercomes in, glances at, is my fridge is running, why is that too warm?And the moment it becomes part of like the standard operating procedures for a store,
32:35it's at that level of visibility where corporate IT is going to know about it.And that to me makes it less forgivable that this happened.No, I can dig that, yeah.I absolutely get what you're sayingNow look just staying on IoT stuffIs oh by the way I went to that websiteAnd I did look and I saw you knowThat's how I saw what the you know form factor the thing was
32:58Their website for that device is all about how they're doing likeOh it's really secure now and it's great and it uses HDTPSAnd it's like oh yeah you guys you knowBut yes they're definitely on the back footBecause all of the messaging on that website is about security nowBut staying with IOT OT stuffBooz Allen Hamilton has just put out a report where they looked at using AI to do sort of O.T. environment hacking and they found that it could pivot around and, you know, do some interesting stuff. I don't think this is particularly crazy research or anything, but it is, I guess, noteworthy when a company like Booz Allen Hamilton is doing these sort of reports. I mean, my joke about this one is that, you know, maybe Booz Allen Hamilton contractors are going to wind up.developing this as a capability for like cyber command or something and we'll all get to find out
33:48about it because it's always their staff who leak all of these secret programs right uh do it pulla pull in a snowdon um now uh this one's great uh Microsoft and uh and friends have disruptedthis huge chinese language fishing kit ecosystemmark let's hear from you on this the store the like this was such a sophisticated marketplaceLike, it's actually really impressive when you look at what this criminal organization built here.It's all been torn down now, but, you know, wow, I've got to say, I was color me impressed.Fishing Marketplace, we love to see it.And you can buy warmed up accounts.
34:30You can buy pre-built login pages for various services.They work very well.And one thing that I found interesting is that how are they selling warmed up accounts?and part of it was that they would go and and buy domains from legit organizations that led the domains expireor maybe they had some you know some companies could go bankrupt and stop but they you can just recover thedomain and and you use that to take over their Google workspace they're you sign into Slackyou log into HR systems because everyone assumes you have ownership of the domain and so that'sbeen a pretty large attack surface. Yeah, and I'm guessing with Slack too, like there's all of those
35:19like federated connected slacks and whatever and it's going to get you into all sorts ofinteresting places. But yeah, it's good Microsoft. We're able to shut that one down. We've gota whiz through the rest of these reasonably quickly. This one's interesting because the headline saysfeds crack down on North Korean IT worker schemes with new raids, charges and arrests. Andand you caught something interesting in this, James,which is the feds are touting that they got nearly 200 laptops during 40 raids.And you're like, well, hang on, that's only five laptops per raid,to which my joke here is, you know, this is less of a laptop farm,
36:01more of a laptop hobby farm by the looks of things, right?But yeah, the numbers here don't quite make sense, do they?They don't make sense from a, like, juice worth of square years.sort of thing, but they do paint an interesting picture of, like, I wonder if this is intentional.Like, do the operations decentralize this for this exact reason, right? If they've got lots andlots of people running five or six laptops, that's going to be so hard to dismantle anddisrupt that scale. It's almost like that's a resilience property built in. Maybe. I mean, I did saya while ago when we started seeing these raids that this was going to be the template for how
36:38the FBI tackled this issue, which is to just find anyone who's probably.oxying those sort of connections.And, you know, so I think these sort of raids, they're just, they're just turning the handleon them these days.It's not like the FBI needs to work particularly hard on this stuff.No offense to the FBI people who are listening.It were like, hey, man.But you know what I mean?
36:57It's not the first time they're doing this, right?So, I don't know, it probably still is worth the squeeze, considering the harm that thatstuff causes.What do we got here?We got a federal bankruptcy court in the United States.its system, I think it was being, yeah, being run by BAE systems.One of their systems got disrupted in some sort of hacks,so that's been causing some drama.
37:22VK in Russia, which is, you know, V-Contact, which is their equivalent of Facebook.That's been having some trouble because, you know,their data center got blown up by the Ukrainian.So this is interesting, right, because we've seen the Yandex stuff getting blown up.Now we're seeing V-Contact stuff get blown up.And, you know, right on Q.404 media published a story saying data centers are war targets now.I mean, I think the Iranians proved that with some of their early attacks against
37:48like Amazon regions earlier this year.But it is an interesting moment, I think, because we are seeing, you know, TomUren wrote an excellent addition of seriously risky business about this last week.There's some irony here in what's happening with Russia because they started hitting Ukrainiandata centers.And Ukraine had already done a bunch of lift and shift into Aziz.here, right? So it didn't really cause like an absolutely massive impact. So Ukraine in responsestarted hitting their data centers and Russia's done a huge data sovereignty push, right? So they're
38:22trying to put all of their stuff in their own data centers. Yandex has only got six.Ukrainians hit three of them and stuff's been falling over left and right. So I do thinkthe Ukrainians have proven that, you know, in a country that where they're trying to reallycentralize their data and do this data sovereignty push, attacking those sort ofcenters can cause real problems. So I do find that interesting. Now last week I also, I wanted tofollow up on this because last week I mentioned that like I had, you know, it felt like a lot ofRussian bots disappeared. When this Yandex thing happened, all of the data since says thatnothing changed. You know, you've got anecdotal reports, though, saying, well, Tucker Carlson's,
39:06you know, Tucker Carlson's engagement on social media dropped by like 70%. We've seen that number.pop up in a few spots.But then we're seeing people who claim to measure this stuff saying there was no difference.James, you had said something on this, which I agree with, which is there's a disconnectbetween the data and the experience here.Yeah, like I'm not a heavy Twitter and Blue Sky user, but I honestly mistook Twitter or X for BlueSky the other day when I was scrolling through just because it was such a materially differentexperience, right?
39:36Normally I go to Blue Sky when I want sensible commentary, it's not full of nonsense.but I was on Twitter and I was looking through the comments and it was relatively sane,which is just not the words you tend to use with an X or a Twitter feed.So look, it's hard, right?Because this is all data points of one with you and I sort of looking at our feeds,but something did happen.The experience is different.The data doesn't stack up.
39:59I mean, if you scroll back a month ago in the White House Twitter feed and scroll it,it's all like people cheering every announcement, every tweet,and now you look at it and it's just like hostile.man. I think the midterms are going to be interesting because it's downright hostile on X,which is a platform, you know, favoured by conservatives in the United States. So it's just interesting.But, you know, Mark, I know you're unconvinced by this. I mean, it certainly feels to Jamesand I that something changed here, but you're a skeptic. I haven't noticed anything, but admittedly,I don't look at the comments under the White House posts. You're missing out, man. That's where you've got to go.
40:35It's what the good stuff is. I don't want to see it.Cloudflare has an interesting blog post analyzing the Andex outage,and they did mention that most of the impact was search indexing that the index runs.They mentioned that the bots just migrated to other ASNs,and there was a small drop in Vaughn scanning activity,but otherwise for bot activity, supposedly it was not noticeable.I mean, I think we are being the tinfoil hatters at the moment here, James,which is an unfamiliar position for us to be in,
41:08but there you go.Now, look, I just mentioned 404their story about data centers.They published an absolute cracking yarnabout a Chinese crime ringthat hacked into a system at Staplesin the United Statesand used their access to this system
41:26to mint 15,000 gift cardsworth over $5 million.And then they sent their mulesto go and buy all sorts of electronics and stufffrom Staples.They sent them back to China to be sold for profit.I just love these really intricate scams that rely on hundreds of mules and whatever.We see it with like ATM cash out scams and whatever.
41:47It's just like it's so much effort to go through.And I'm sure they made money, but like, wow, that's a lot of work involved in that.What else?Microsoft has dropped patched Tuesday, 1400 bugs.I just figured it was worth mentioning the number.I wonder what sort of functionality it's going to break because it feels likewe get patch Tuesday now and then the week after the patch Tuesday news we get all of the news aboutlike oh people can't use their clipboard if they're running this software and like this thing
42:17randomly crashes and just the just the QA on patches is just got it's got too hard to bothclose the bugs and make sure that they're perfectly QA'd not throwing shade at Microsoft on thatI think everybody's in the same boat uh GitHub is expanding its malware advisories toeight package ecosystems.It was just doing this for NPM, right?So, James, talk us through this one.Yeah, so this is, it's come out from the Dependibot team within GitHub, and I put thisunder the category of good to see, right?
42:49Dependipod is a useful service.I get PRs every now and again, regularly of things that need to be updated, and I probablywouldn't pay attention to them if it weren't for Dependobots.So it's a good service.Extending that to the other package ecosystems is great, good to see it.But it's just, I don't think this materially changes the frequency or the, you know, the problems being caused by supply chain attacks.Because by the time a Dependabot, PR reaches your repo to tell you that there's a malicious package in a given ecosystem, it's too late, right?Because they're getting picked up in the developer's workstation or in the CI workflow.
43:22So good, but don't rely on this to put a big dent in quelling supply chain attacks.Yeah.Now, speaking of supply chain attacks, and I want to get your thoughts on this one.Mark, this looks like a pretty cool sort of supply chain-ish type of attack going through a connect-wisebackup app into Microsoft 365 tenants.And we all know once you're in a Microsoft 365 tenant, you know, you can actually swimdownstream into, you know, an actual Windows network.So it's just like good hacking here.
43:55But walk us through this one, mate.Like, how did they do this?Yeah.So if you're an IT team and you want to back up like mailboxes, data on OneDrive.and SharePoint, you can use tools, and this one connects to your cloud, has breed access to all the data,also has right access to be able to recover the data. So inherently, it's a tool that's in a privileged position.Now, folks found that there was a login through that app that's installed, but the login didn't comefrom one of Azure's IP addresses, came from an external one. That's the first red flag.
44:30and gaining access to the data was one issue,but in many environments they found that the service gave itselfa global administrator and was able to add new credentialsand access other applications on the tenant.That's a bit harder to explain.Don't think it generally needs this level of global admin on the tenant,but also in some case it was just like an old servicethat's no longer used.
45:00So another just problem of not cleaning up things correctly.And one last interesting point,I found that the Hunter's team has a blog post on this,and they said that some of the backdoors were attemptedvia Microsoft Graph or MS Cloud API,and the Microsoft Graph directory reads leave no audit records.So that's another thing that makes it harder to respond to.Yeah, yeah, that's not helpful, we would say, right?
45:31So what I'm guessing someone managed to grab, like, the authentication tokens or something from ConnectWise,is that roughly how this would have gone, James?Yeah, it's not your O-Orth, as I understand it, this is like statically provisioned service principal accountswould have had an API token that was handed off to ConnectWise.But similar to Mark, when I looked at the reporting where it says, you know, the cloud managementMS teams granted itself, granted itself, global administrator.Yeah.an entry expert, but I don't think that's a good idea. Yeah, well, now I'm king of Australia now,
46:03because I said so. Seems a bit like, yeah. I mean, it is a little bit of a blink moment when yousee that in the writer. There's over-provisioned, but then there's self-over-provisioning. That seemsto have gone too far. Yeah. Now, we've got a report here about adfaction. This is a Cyberscoopwrite-up on some work out of Citizen Lab. And look, I found it really interesting. Basically,they're looking at this Israeli and, you know, digital advertising company that looks like it managesto squeeze malicious ads into the major ad distribution networks like Google and whatever,which is, you know, cool and, you know, so far so normal, right? Like if you, sorry, I shouldtake a step back for a moment, right? They are squeezing into these, into these ad networks in order
46:48to deliver, you know, iOS malware to specific targets, right? So it's like a targeting platform.That's the allegation here is that this advertising network is used to distribute, you know, malware like Kandaru in a targeted way onto particular devices.So look, that part of it's really not surprising.What I do find interesting here, though, is that there is apparently this angle to it where this ad network looks like it is trying to de-anonymize people as well,and sort of tie their online activity to identify as like, you know, in selectors like phone numbers and whatever.I find that more concerning, actually, than a malware delivery network existing like this,because it's just something that's so totally expected.But the idea that SIGN agencies or state-backed actors or their contractors are starting to really infiltrate the ad ecosystem in a widespread way to do, really, let's be honest, a form of mass surveillance,
47:48that strikes me as being much more concerning than the malware delivery vector.James, what was your feeling on this one?100% agree.I mean, this is an advertising industry,an industry that has a huge amount of money running through itthat is wholly and solely geared towards knowing,understanding, triangulating,and predicting how to get the right message
48:07in front of the right person at the right time.And that's in the good case.But when that message is a link to a watering holethat has an exploit,it's no different.The technology can be used in all the same waysand all the data that they collect to do advertisingin a way that a consumer might expect
48:20is just as equally valuable to an advertiseryour intelligence operator.It's real funny, man.I remember when the Snowden leaks first happenedtalking to someone in the intelligence communitywho was big mad about the leaks being misrepresented at that time.Now, this was very shortly after the leaksand just something they said to me struck me.
48:41It stayed with me.If we could do like a hundredth of what Google doesto violate people's privacy,we would just have incredible insights.Like, you know, we are careful, blah, blah, blah, blah, blah.That was the gist of what they said.But, you know, that's what's crazy about this is like, someone here is having a crack at like,okay, let's be Google.
49:01And that's, I wonder, you know, it doesn't feel like something that would pass legal master in the US.I don't know.But it's, look, it's an interesting story.It's linked through in this week's show notes.People can go and read it.Just going to run through these quickly.There is a Yvanti bug being exploited in the wild.It is a chain of three bugs.
49:20Sisser is warning about that.one. We also have a Beyond Trust IT support appliance O'Day being being exploited since August.It's being used by a group called Armory Castle. They apparently really know these Beyond Trustappliances quite well. Mandiott has seen them use custom malware for it and stuff. It's,you know, pretty interesting attack. But like, yeah, really like a Beyond Trust remote accessappliance like that, just those words send to shiver up my spine. You're expectingproblems immediately just with that sentence. I mentioned it earlier. Drupal ispatched a critical RCE. Palo Alto Networks has just fixed a pan-OS bug that's
50:03being exploited for root-level code execution sounds like a good time except forpeople who run pan technology. Eight US states are suing open AI over allegedharms to users. What I found interesting about that is I looked them up and allnone of these states are democratic states, right? So one is Arizona which is kind of apurple state. I just think it's really interesting that you've got this Republican presidentwho's very pro-AI and then you've got eight states that lean heavily Republican trying to sue open AI,right? So the politics around AI at the moment are just fascinating to me. There's been some sortof attack too that has halted work on a Denmark to Germany tunnel. I only mention it because we love
50:45real world impacts. That's one for the slide decks for all those people preparing their powerpoints outthere. Yeah, Ransaware groups are expanding their targeting of a Palo Alto network's vulnerability.Just, just crazy. So much going on. But the last thing I wanted to mention, too, is, you know,there's been this ongoing story about some activist in the United States who had his phoneseized when he was crossing the border into the US. He's actually an American and had hisdevice, they asked him for his device pass code and he put in a duress password. And he was runninggraphene so it nuked his device and he's being charged for that and could go to jail for five years.There's just been some interesting developments in this and one of them is that they
51:27they had actually marked him for device seizure because they knew he was crossing the border.So this to me says that these permissions that US authorities have to examine people's devicesas they cross a border are being abused because if you want to look at somebody's deviceyou should get a warrant.That's how this works.You shouldn't just abuse a.customs authority to get a warrant-free look on someone's device.This happens regularly and it's bad and I really hope that prosecuting him for this
52:00backfires and we wind up with a precedent that cuts the other way.But again, people who want to check that one out can read it in this week's show notes.But that is it for this week's news segment.I think we've gone reasonably over time.It was a good time.James Wilson, Mark Coory, thank you very much for joining me totalk through all of that. It was very, very interesting.Thank you. Thanks, Pat. What a week.
52:30That was James Wilson and Mark Coory there with a check of the week's security news.Big thanks to them for that. It is time for this week's sponsor interview now, and thisweek's show is brought to you by Trail of Bits. And in this week's sponsor interview,Trail of Bits principal security engineer, Anish Naik, joined James Wilson to talk through a bunchof work that Trail of Bits has done in cooperation with O'Shaeek.Open AI. They are working together on the so-called Patch the Planet initiative. And they havereleased two open source plugins, patching bugs and regression hunt. And that is what Anishnaicjoined James Wilson to talk about. Enjoy. We're going to be releasing two plugins. One of them is
53:18called patching bugs and the other one is called regression hunt. Now, did marketing get involvedin creating these names?No, it's all me.That sounds a lot like engineering, yeah.Yeah, yeah, this is exactly what it does.Yeah, we're literal people here.And, yeah, it's effectively an amalgamation of what we've learned through over 60 differentprojects on Patch to Planet.
53:42And it's effectively looking at all the different PRs, every single PR that we've mergedinto an upstream repository, as well as looking through thousands of codec sessions ofthe engineers that were on these projects.And we said, okay, well, can we condense and distill this information into best practicesthat we can then provide to open source maintainers to deal with, you know, all the bugsthat they are already facing and probably will continue to face in their different projects.Give us a sense of where does a developer actually use these.I mean, we've got the two plugins, patching bugs and regression hunt.
54:18But where in the life cycle of actually fixing all these delugees of bugs do these actually get,by a developer?I wanted to be as easy as you have a bunch of GitHub issues that have been opened by peopleout there looking to use your product or found bugs.And I'm hoping that people can just take these plugins, provide a list of PRs or commits oreven issues, and say, is this bug legitimate?If yes, come up with a minimal patch that solves the violated security invariant.most importantly, let's check our homework. Let's make sure that we didn't introduce a bigger issue
54:56or like an equally bad issue, again, all from a security perspective. We want to make sure thatwhatever patch goes through does not make the system worse than, you know, before we started withthe patch in the first place. And so that's sort of how I would like for people to use it. You just,you throw them in, say, look at this commit and then it goes from there. Which sounds, you know,the framing there's good, but you've almost oversimplified it in terms of why.this is such a problem because, you know, just even, you mentioned sort of three steps there.Is this bug worth fixing? How do we do the minimal amount of fix to fix it? And then how do wemake sure we haven't broken things? Those are not, I guess, I would say, characteristics that are
55:36inherent to coding agents as we see them today, right? Codex and particularly Codex wants to solve allthe problems all the time and feels good when it's solved far more than you wanted to do.And so just even the notion of knowing what to fix is not necessarily something that a model can do a good job with unless you nudge it in the right direction.How do you go about telling a model almost like calm down, don't fix everything, don't fix every single problem you see, just do that initial triage.How do you nudge the model in that direction?Yeah, that's a great question.So the first thing I'll say is, you know, I assume that the agent will get better.If I asked an agent to patch last year versus this year, Codex objectively has figured out that
56:24patching is not just writing the diff, but it's also testing it, not with just a unit test,but also let's try to do an integration test.So it knows that there's more to it.But, you know, so I'm going to assume that it gets better.And so what I wanted to do through this process is ask what are the gaps between what, you know,are Trailobits engineers are doing.Like if I were to do it manually, these are the steps I'm.would take. And so we looked again at the session data and the PR data and said, okay, well,
56:52it's really good at this stuff. But if I could just tell it to do a POC, I want user interaction.I want source to sync, right? I want to see where user flow data is going. And so, you know,the little nudges are use a POC. And so the entire, all the plugins are driven by live reproductionsof the underlying bug or vulnerability on the actual system.So it's not just a unit test, oh, I, you know, like a unit test to be like, oh, there's a bug,and I have this unit test that shows the bug.It's like, here's the bug, and I can show you how to get there.Right.
57:29Through on a live deployment.And then the second component is with regression specifically, right?Like, we don't want to introduce more issues is it's more important to tell the agent what aregression is not than what it actually is.Because a regression can be a thousand different things.It could be a built compatibility issue, a performance consideration.But what it's not is more important because that reduces that noise to help defenders at scale.Right.
58:00Like I want to, if I have a patch, I want to know quickly there's no issues or there's maybe one versus like 10 possible ones.Well, now I went from one patch to 10 possible new bucks, right?that's what that's that's that's like the antithesis of the goal and you can always find more regressionsif you broaden the scope of what is a regression and so i guess it's like yeah it's it's knowinghow to define the regressions we care about i i guess is the important thing but you know thatwhat you mentioned there in terms of the patching bugs side of this and we should say these these arebeing released as open source plugins and and basically as i understand that these are skill files isis that the right way to think about what this is it's not like a harness or something some new tool or it's
58:40going to integrate directly into what you're using? That is correct. They're very lightweight skillfiles. And so, yeah, each plugin is like, I think, five to seven skills. And you can pick and choosewhich ones you want. So let's say you have a patching pipeline you're really happy with. You can justuse parts of the regression hunter and vice versa. So yes, they're just, they're skill files.Okay. And so, you know, the patching bugs work, right? Do the triage. Is this worth fixing? Come up with theminimal fix. I can see how a skill is super useful there because it does, it governs these modelsthat increasingly just want to run a muck and give you, you know, 1,200 solutions for 500 bugs.But, you know, I want to push back a little bit on the notion of the regression hunting aspect
59:24of this, because, you know, we've got this ridiculous expectation that software should be stable,right? And yet, it's so complex. And the way in which we have sought to improve the quality of softwareover time is with a lot of deterministic testing,unit testing, functional testing, integration testing,end-do-end testing.Why would I want to throw a non-deterministic large-language modelinto the process of ensuring I don't have regressionswhen really what I'm asking for is deterministically prove to me
59:51that my software will continue to operate the way it always has,but now with this bug fixed, right?Why add an LLM into this for regression testing?Totally.I think I'll answer it in a few different components.The first one is at the scale that we're dealing with bug reports for these different projects,you kind of have to meet that scale with something that also scales really well.And so like AI becomes this almost like natural provider.
01:00:18And more importantly, I think like there's a lot of effort that goes into using AI for adversarialwork, building harnesses and all that stuff.But like, you know, it's good to find bugs.You want to find the bugs, but it's only good if you can like also fix them.And so I think a lot of the work that we do is try toarm defenders with this, you know, this solution to use it effectively. So I, you know, and I don't thinkit is, you know, the final, you know, the AI is going to perfectly, you know, fix your patches.And maybe it will get there one day. But it's sort of like another kind of step in the middle of,
01:00:52okay, you have these, you know, 20 bugs and you have 20 patches and the, you know, with the plug-insand, you know, hopefully you have some kind of a threat model or your own opinions about what yourpatches look like, you have a prioritization exercise of saying, hey, James, here's these 20patches. I think you should fix probably these two or three because they're reachable from theuser perspective. And so it's basically part of a prioritization exercise of you're dealing withso many. How do we like get to something that's actionable quickly? And how do I like empower opensource maintainers that don't have, you know, a whole army or security team? Like it's two orthree people who are just passionate it. Like how do they how do they deal with the hundred bugs, you know,
01:01:31when they have 100,000 users across the globe, right?So it's kind of, it's a step in that direction.Yeah.And you mentioned also that this is sort of a,there's a point in time aspect to this as well, right?Where last year, if you asked a model to patch a bug,you'd get a very different behavior and outcome.Today, I find, you know, Codex in particular is like,
01:01:51I fixed that bug and I found these three things along the way.Do you want me to fix those?And you say, yeah, go fix those.And then for each one of those, it's like, I fixed it.And I found five things along the way.And it's like, you can sit there and use,your entire token budget over and over again, just saying, yeah, okay, go fix all the things.And then your net result might be some, you know, god-awful mess.
01:02:11But that's just the way they're operating right now.As you said, they're going to get better.How do you think about evolving these skills?Like, is this just something, TrailerBits has said, look, we've learned this, here's somethingthat's useful have at it, or do you want to keep evolving these and changing these over time?That's a great question.I think to the first point, I think a lot of it comes now to opinions and take.like how you guide these agents, like I'm always going to assume they're going to get better.
01:02:35And it's about like guiding them and telling them like, this is what I want it to look like.So this is what I consider aggression.This is my threat model for my system.These are my patching guidelines.And so if you provide the opinion, you can guide the agent the way you wanted to outside of whatever it does in its own best practices.And then the most important thing for us is these plugins are live things.you know, we already use these every day in our, you know, our new wave of patch the planetprojects.
01:03:04And I am actively collecting all that trace data from all the engineers.And, you know, I would love to get feedback for maintainers and open source projectowners.And I want to collect all that.And I want to make it better because I do think that like, you know, defense of, you know,patching is not the most fun thing.Obviously, it's better to find RCE.But like, you know, you got to, someone's got to fix them.
01:03:23And I think fixing them is more important.And it's, yeah.So, yeah, we're hoping to have a really good feedback loop where these plugins get better over time.All right.All of this makes a lot of sense.I'm actually excited to try it out myself.So where do folks find this?How do they get on board?
01:03:39For sure.So these plugins are open source public.You can find them at github.com slash trail of bits slash skills.Outside of these plugins, we also have obviously a lot more.And again, we do really care about this a lot.Please provide your feedback as a poll request, as an issue.I'm going to try to incorporate it, you know, as much as quickly as possible to make sure that, you know,everyone in the open source world has access to try to deal with, you know, this pertinent problem
01:04:09effectively.All right.It sounds awesome, mate.Thank you so much for dropping by for a chat.Awesome.Thanks, James.Appreciate it.That was Anish Naik of Trail of Bits there, joining James Wilson to talk about a bunch of the work
01:04:23that they've been doing with Open AI.We're going to wrap it up there, though.it for this week's show. I do hope you enjoyed it. I'll be back soon with more securitynews and analysis, but until then, I've been Patrick Gray. Thanks for listening.
There aren't comments yet for this episode. Click any sentence in the transcript to leave one.