AiCyber.Land #55 - OpenAI's Linux Bug Hunt & Safer Coding Agents

Show Notes

Sources & transcript: Episode #55 page
Read transcript: https://share.transistor.fm/s/c3e4f779/transcript
Apple Podcasts web strips note links: use its existing Episode Webpage action for clickable sources and the transcript.
Automated transcript; may contain errors.

What happens when AI can hunt for Linux vulnerabilities at scale—and your own coding agent can reach more files and credentials than you intended?

Bryce Kunz and Shelby Pettit discuss OpenAI's reported Linux-kernel research: moving from candidate findings to working local privilege-escalation exploits, generating candidate patches, and the review work that still belongs to maintainers. They ask what increasingly capable exploit-finding agents could mean for defenders and software teams.

Then Shelby walks through a Wiz security checklist for Claude Code: prompt and data egress, generated-code quality, hallucinated package dependencies, access control and secrets, and autonomous permission modes. The conversation covers slopsquatting, limiting an agent's blast radius, pre-tool hooks, local redaction workflows, and why AI-assisted reviews should complement—not replace—traditional security controls.

Plus: small local models, Bryce's plans for recurring code and server-configuration reviews, Shelby's Harry Potter puzzle, and a pre-watch Godzilla take.

SOURCE CONTEXT
OpenAI's November 4 report distinguishes 2,308 candidate findings, 1,724 reproduced vulnerabilities and 948 unique local privilege escalations with working exploits in its tested configuration.[17] The reported cost is $1,213,430 in API credits—not a count of tokens.[17] Candidate patches do not mean every issue is already fixed or that all Linux configurations are affected.[17]

The linked Wiz resource is its public cheat-sheet landing page; the full PDF is form-gated and was not independently reviewed for these notes.[50] The original huggingface-cli experiment was reported by Lasso Security on March 28, 2024: an empty package received more than 30,000 reported authentic downloads in three months, not 30,000 confirmed compromises.[60] The specific 2023 timing recalled on-air is not established by that report.[60]

Find the show: https://aicyber.land

KEY MOMENTS
0:00 Welcome, sweet soup & hype-man ambitions
1:04 OpenAI's reported Linux-kernel bug hunt
2:43 Local privilege escalation & container risks
4:18 From candidate bugs to exploit chains & patches
6:42 Novel techniques, disclosure & the future of exploit research
9:17 Wiz's Claude Code security checklist
10:03 Prompt egress, sensitive data & exposed credentials
11:52 Generated-code quality & security constraints
12:54 Slopsquatting: hallucinated package dependencies
13:43 Agent identities, access control & blast radius
14:27 Autonomous modes & sandbox boundaries
15:04 Pre-tool hooks—and their limits
16:32 Local redaction workflows & small models
18:04 Recurring code reviews, server checks & defense in depth
20:10 A Harry Potter puzzle with a fourth dimension?
20:56 Godzilla plans & pre-watch expectations
21:26 Wrapping up

What boundaries do you put around your coding agents? Share your approach, and subscribe for more AI and cybersecurity conversations.

Sources

[17] https://openai.com/index/a-thousand-fewer-paths-to-root — A thousand fewer paths to root
[50] https://www.wiz.io/lp/claude-code-security-best-practices-cheat-sheet?sfcid=701Vh00000gJQZuIAO — Claude Code Security Best Practices Cheat Sheet — Wiz public landing page
[60] https://www.lasso.security/blog/ai-package-hallucinations — Diving Deeper into AI Package Hallucinations — Bar Lanyado / Lasso Security

What is AiCyber.Land?

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.

Episode 55 — Ai Cyber .Land
AUTOMATED TRANSCRIPT — LIGHTLY REVIEWED; ERRORS MAY REMAIN

This transcript was generated by Whisper medium speech recognition and lightly reviewed by an AI assistant using the complete ASR text, existing chapter/uncertainty notes, known spellings, and saved timing evidence. It was not human-transcribed or verified by listening to the audio. Recognition errors, omissions, and incorrect names, model names, numbers, or phrasing may remain. This is not a fact-checked record: on-air/ASR factual and numerical claims have not been silently rewritten.

Notation: [unclear name: ASR "..."] and similar [unclear term/model name/words/comparison target: ASR "..."] labels retain uncertain automatic wording without guessing a replacement. These labels and this introduction are editorial notes, not spoken words. Broken or incomplete ASR phrasing is otherwise left in place; no missing speech has been invented.

Timing: [m:ss] labels are the original approximate starts on the FINAL edited-video timeline. They are preserved unchanged; no second speed/intro transformation has been applied. Existing timing evidence was checked offline; the published/hosted audio was not accessed. This TXT is not an SRT/VTT or a claim of synchronized Apple/Spotify transcript support.

Hosts (editorial credit): Shelby Pettit and Bryce Kunz.

TRANSCRIPT

[0:08] Hey, welcome back to the Ai Cyber .Land podcast. This is where artificial intelligence, super
[0:13] intelligence, and some other stuff combine together and they make a sweet, sweet soup.
[0:18] A soup sweet, do you think? I don't know. That's a stretch on my part. But you got me,
[0:24] Bryce, a human, and you got the world's greatest co-host, Shelby. So we're back again to be
[0:30] AI in cybersecurity related. The breaches, the breakthroughs, and everything else in
[0:33] between. All right, so grab a drink, kick back, and smash that subscribe button. You
[0:38] knew it. You knew I was going to say it because I always do. You might even, if you got some
[0:41] extra hypes, give us some hypes. I don't know what a hype is, but I heard it's good. So
[0:46] and let's just be honest, my dream in life is just to like be a hype man, you know, be
[0:51] somebody's number two where I just come in and get people jazzed, you know, so. That's
[0:54] like the secret power because you have a lot of influence as a hype man. But none of the
[0:58] responsibility. Boom. Boom. Career goal obtained, locked in. Well, this week we saw OpenAI lock
[1:08] in and specifically they set their sights on finding vulnerabilities in the Linux kernel,
[1:15] which resulted in all these people back here losing their minds. But the cool part is they
[1:22] took their latest model that has not been released yet. This is like a, you know, one
[1:26] that's only in the labs and they spent what would have cost them or would have cost a
[1:30] normal human over a million dollars on tokens and said, hey, you have this budget, go find
[1:37] as many vulnerabilities as you can in the Linux kernel. And for those of you that don't
[1:41] know, Linux is what powers most of the web. So like when you go to a website, it's typically
[1:46] like a Linux server on the other side, right? So you use Windows or Mac, but people that
[1:50] provide the service to you typically use like a Linux OS. And so a lot of the internet is
[1:56] powered by Linux, right? Especially on the production side. And so the AI, it found about
[2:03] 2300 candidate vulnerabilities. And of that, it was able to whittle that down into about
[2:09] 950 working exploits. So almost a thousand working exploits. Now typically Linux maintainers
[2:18] would lose their mind over one exploit, right? Or like have a bad week over it. And these
[2:24] individuals got almost a thousand working exploits. Now we don't have any of the details
[2:28] about the exploits. So we don't know if they are like affecting all versions of Linux or
[2:34] if they're affecting the default configurations of the Linux kernel. But due to the sheer
[2:38] number that we're talking here, we have to assume some portion of those are going to
[2:42] affect everybody who's running Linux that is out there. These are considered local privilege
[2:47] escalation or LPE is like the acronym exploits. So essentially if you, you need access to
[2:53] the server to begin with, like you need like a user on the computer. And then when you
[2:57] have that, you can privesc to root. The reason why that's a big deal is a lot of times when
[3:02] people build web apps, they put them in what's called a container. And you can imagine a
[3:05] container is kind of like a, like a shipping container, right? It has the app inside of
[3:09] it. And they put those containers kind of like a boat and then the boat like is kind
[3:13] of the Linux OS, right? And then you are accessing the app that's inside the container. And a
[3:18] lot of times these local privilege escalation exploits, another subset of those will also
[3:23] function effectively as like a container breakout. So they'll enable, if someone finds a vulnerability
[3:28] in your web app, it will enable them to get root not only on that specific app, but also
[3:34] on all the other apps that are like co-hosted on that same physical hardware. And because
[3:37] people want to be, you know, hardware is expensive, people want to be efficient with it. So they
[3:40] typically run a lot of different containers on the same hardware. In fact, there's entire
[3:45] like cloud services, like if you've ever heard of Kubernetes or things like that, which are
[3:48] really just designed to maximize the amount of containers you can run on hardware. So
[3:59] zero-days without you really doing anything. I mean, I think we've seen repeatedly the
[4:02] answer to that is yes. It was nice that they like put this on something that was for those
[4:06] of you who don't know Linux is typically like a free OS. It's like open source. Anybody
[4:09] can download it and run it. That's why it got so popular in this commercial space because
[4:13] it cost you money to run it and scale it. But the cool part is like, typically there's
[4:18] like a multi part to finding vulnerabilities. There's like a vulnerability discovery phase
[4:22] where you find potential bugs in the code. Then you have to like validate those bugs to
[4:27] like make sure they're going to be applicable to whatever use cases you're currently working.
[4:31] Then a subset of those are going to be actually exploitable. So there's like these
[4:36] primitives to exploits that you have to get working in modern operating systems. And if you
[4:40] can't chain enough, chain together enough of the primitives, then you can't get the full
[4:43] exploit chain working. So even though like you might have like a vulnerability and it's a
[4:46] vulnerability here, if you can't chain together several vulnerabilities around it, you can't get
[4:50] the exploit working. Like you can't break out of the container effectively. But then they have the
[4:54] AI automatically chain all that stuff. And then they even went and they generated patches for
[4:59] the code. And then they gave all that information back to the maintainers of the Linux kernel so
[5:04] that they could hopefully roll the patches in. Now, that is better than most researchers do.
[5:10] Most researchers don't really generate the patches. So that was super nice of them to do,
[5:14] in my opinion, and a good thing that everyone who's using AI should be doing if you're reporting a
[5:17] vulnerability, especially if you have access to the source code, which you would in an open source
[5:21] project. But still, like as a maintainer of the project, you'd have to like vet it, make sure it's
[5:25] not going to like you have to put it through like regression testing to make sure it's not going to
[5:28] break anything else. So there's still a lot of work for them to do. But OpenAI is basically saying
[5:34] like, hey, we gave them all the information, including the patches, and they're just going
[5:36] to slowly work on fixing all these vulns, which honestly, like in the Linux maintainers didn't
[5:42] have to pay for credits, and they didn't have to do any of the work other than now they have to
[5:45] go fix all this stuff. So I don't know, I thought that was cool. I thought, you know, it'd be nice
[5:50] if there was more stuff like this, especially the patch generation step. Yeah, that would be
[5:55] really helpful. And yeah, I, you know, I think unfortunately, if you are an exploit developer,
[6:03] and that was your career path in life, you know, you're now in your new career path is going to be
[6:09] tell OpenAI or [unclear term: ASR "cloud"] to do your job. Because like there's no way you're producing a thousand in that
[6:13] same time period. But then again, I mean, they also did burn over a million dollars in tokens,
[6:17] which, you know, the cost of that will go down over time, right. And you could also say they're
[6:21] using a frontier model nobody has access to. But typically, what is open sourced in the LLM space
[6:27] is about six to nine months ahead of what the labs have. So, you know, in another year from now,
[6:31] this same technology will be in every cyber criminals hand that exists on the planet.
[6:35] What's your thoughts about that? How do you feel about that, Shelby?
[6:37] Oh, nothing to see here.
[6:42] I follow [unclear name: ASR "Clint Gibbler"] and have for a little while and he's one of the researchers who's been
[6:47] working for months on this. And it was interesting to read kind of his private thoughts on this,
[6:51] because he's like, I've been working on this blog post. And then it was this one, right.
[6:55] And hearing his thoughts about it, I don't know if you mentioned this, but there were 10 novel
[6:58] Linux exploitation techniques that they discovered, like just brand new. And then I
[7:04] guess the guy, [unclear name: ASR "Kyle Zang"], who built the pipeline for this work and then did most of the research,
[7:11] he described him as like bouncing around like a little kid at Christmas every time he found
[7:14] another few hundred zero days. Like, like, I just can't. It's so crazy that like zero days have
[7:20] become so accessible. I guess you still need a million, $1.2 million to get it. But
[7:26] yeah, it's kind of cool. And how they got like desensitized. They'd be like, oh, this one's just
[7:32] like an unauthenticated RCE, but you have to like click once or twice. Yeah, not that big a deal.
[7:37] Like, they're talking about how like the bugs they're finding are kind of severe and how like
[7:42] a bug that is so impressive that would have been the talk of the year, they're now finding on a
[7:47] weekly basis. And so that's, it's wild. I love to see the responsible disclosure and like
[7:53] helping out the maintainers by actually making the patches is wild. And I like that as it,
[7:57] I would love to see that as a new trend. I think it just goes to show anything that you think is
[8:07] every day, like you rapidly become apathetic to it. You're like, you're like, ah, another giant
[8:12] diamond. Get out of here. You're like, I gotta, I gotta mine more of these diamonds.
[8:16] I know, just like zero days, zero days, zero days.
[8:20] Coming from someone who's definitely been in that. Yeah. I, you know, I think you just get apathetic
[8:25] to wherever you're at, right in life. So I, and whatever you, whatever's normal for you.
[8:30] Um, yeah, it'll be interesting. Like I wonder, is there like a, like, obviously there's a lot of
[8:36] zero-days being mined out right now, but you know, is there some point where we plateau, like we get
[8:40] basically all the ones we can out of the code bases with AI and then, and then there's like the
[8:45] next generation or do we actually hit like some singularity type point where it's just like,
[8:50] well, whoever can spend the most money on tokens wins, you know, like I, I don't know.
[8:54] It'll be interesting to see what happens. Part of me feels like it's got to taper off at some point
[8:59] because maybe it's like, you're just getting rid of like all the old, like the vulns built pre-AI,
[9:03] but then at some point there'll be like this new set of vulns that are, I don't know, post singularity
[9:08] or something. I don't know. Anyways, we'll see what happens in the future. What'd you see this week?
[9:11] That was cool. Shelby. Yeah. That's interesting to think about. I was looking at a, like a cheat
[9:17] sheet that was published from Wiz and it's all about like security best practices for Claude Code.
[9:24] So I thought it was a great resource. We'll be sure to include a link here,
[9:27] but basically it kind of starts off by introducing you to the idea that like, which is not new,
[9:31] but we know that like, because coding agents have so much access, you kind of have to think of them
[9:36] like a human developer because when you're thinking about like access control and security
[9:42] protections and safeguards that you need to put in place, it's kind of like having a potential
[9:46] insider threat, right? Yeah, for sure. That can do things autonomously. However, they also mentioned
[9:51] that there are five specific risk surfaces they identified that are specific to AI that we need to
[9:57] make sure that we are taking active steps toward mitigating. So the first one is prompt and data
[10:03] egress. So it gave, I mean, very, very common sense things, but like if you're trying to make
[10:08] a best practices or like worksheet for your company and your developers or things like that, like it's
[10:14] this is a really good place to start. It says like, don't put your secrets in there. Don't put your
[10:17] PII in there. Don't put PHI in there. No internal architecture, like revealing where the trust
[10:22] points are in your network. No tokens, no session cookies, things like that. And their guiding
[10:29] principle was would I publish this on like public's Stack Overflow? If yes, go ahead. If not, you
[10:35] shouldn't be telling it to the agent because you have to understand that prompts and all the data
[10:41] that's like made and analyzed, potentially analyzed, is also being sent to the underlying
[10:45] model provider. So once it's kind of sent out, you have to consider it compromised almost. Like if
[10:50] your password is included in something that you sent to in a prompt or other data, then you've
[10:56] got to go rotate those creds. So that one is great. And there are some action items like redact logs,
[11:02] if you're troubleshooting or other things like that before you have AI look at it. And use
[11:07] environmental, environment specific repos and credentials. Another point that it made was
[11:14] similar to what you talked about last time where you said you don't know how your agent got your
[11:19] credit card number, but it did. So it can easily sprawl to look at resources on your network or
[11:26] laptop or whatever or your server that you didn't explicitly tell it to but it's trying to accomplish
[11:30] its task. It's going to be nosy. It's going to look all over the place. So if you want it to not
[11:33] be looking at specific things, well, you can take certain actions. So for example, like so you can
[11:39] put in your CLAUDE.md file like specific files to avoid so that it's not slurping in like
[11:47] credentials or things like that without you even realizing it. Anyway, so point number two is
[11:52] generated code quality. So it gave a helpful like short list of some common code vulnerabilities
[12:01] that are like that commonly happen when you have AI generating the code that which also fall in the
[12:07] category of legacy scanners don't catch it very early. So I thought that was helpful. It gave like
[12:11] very specific heuristics and things to look for. It also mentions explicitly including security
[12:18] constraints in your prompts. Oh, I think I misspoke earlier. That's what you want to put in your
[12:22] CLAUDE.md file so that every time you fire up, it's going to always read that. So like you don't
[12:26] want to tell it just, you know, build a login mechanism, right? That's not enough of a prompt.
[12:31] You want to build in things that you wanted to keep in mind about that login or that page or
[12:37] whatever it is, that resource so that it is knowing to be security minded as it accomplishes its task
[12:44] instead of finding the path of least resistance. And so you want to have these security constraints
[12:49] in your prompts and just put that as part of your startup files so that it's kind of default
[12:54] behavior. Number three, I liked this word. It's one I hadn't heard before, slopsquatting.
[12:59] Okay. I really like that. Basically. So this is when threat actors act on package names that AI
[13:07] hallucinates. So this is all about securing your dependencies. There was a guy in 2023, which in
[13:14] terms is like eons ago, but still I think the point stands, he noticed that AI kept hallucinating
[13:20] this package that was called huggingface-cli. And that did not exist. So he made one.
[13:26] And within three months, it was downloaded 30,000 times. So yeah, just kind of, so that's what
[13:34] slopsquatting is. And just, you want to take steps to make sure that you're, you know, where
[13:38] your dependencies are coming from and that they're real. Number four, risk surface number four
[13:43] is access control and secrets management for agentic workflows. Like, do you know the scope?
[13:51] Do you know the blast radius of your agent? That takes a little work to map out or to
[13:57] deliberately kind of pen in to a space so it can't see every resource on your network, right? You
[14:03] usually want them to work under certain constraints as a best practice. So you want to constrain its
[14:08] identity. You want to keep an eye on those logs, see what your AI is doing. Cause that's what we've
[14:11] seen this year. Sometimes they go off, they go off base and go attack people, which hopefully none of
[14:17] your AIs are doing that, but yeah, it's good to keep an eye on those. And also don't give secrets
[14:21] to your agent, which I mentioned earlier. Okay. Last point it mentioned for the risk surfaces is
[14:27] agentic and autonomous mode risks. So learn the permission modes and how they work in your agent.
[14:34] And then for like the YOLO mode or the bypass permission mode should only be used in a sandbox.
[14:41] At least that's what I was told.
[14:44] What if my whole life is one big sandbox, Shelby? Everything I touch is a sandbox.
[14:49] I was like, I don't know if I know what a sandbox means, but I told Claude it's cool. Don't worry.
[14:55] YOLO it. I said Claude, hey, [unclear name: ASR "Daria"] removed dash dash YOLO. Can you just add that back in for me?
[15:00] Okay. Thanks. Bye.
[15:04] So, um, hey, on the serious side, uh, like, uh, it's a, it's a little more technically complex,
[15:12] but you could tell Claude to just do it for you, which is, uh, when you're using like a Claude
[15:15] type tool, they have like a hook feature and you can actually set up like a pre tool hook call.
[15:20] And so one of the few security things I do, which I know is going to shock people,
[15:25] is I tell Claude to do a pre tool call check that says, is the file it's trying to read the .env
[15:31] file, which is typically where my secrets are. And if so deny that request. So, um, because the
[15:37] difference between the, the, the hooks are more like programmatic calls. So like the LLM can't
[15:43] bypass that. Now you can look at that and say like, that was a jerk move of Bryce. I'm going to
[15:49] write a script and that script is going to import [unclear term: ASR "dot ENV"] automatically. And then I'm going to
[15:52] export the things out of that. So there are definitely ways you could get around it, but
[15:56] you know, like it has to go more off course, let's go more rogue. There's like a rogue meter,
[16:00] you know, and you want to keep it low when your agents are running medium. They're being kind of
[16:08] like, what are you doing? That's my goal.
[16:12] Also imagine it. You have a high risk and you just have all your tokens get eaten up.
[16:16] You're like, huh, what's it been doing? What's it been doing? I don't know. It wrote a root
[16:21] kit and now it tells me I have a ton of tokens again. No, um, the, um, why is it just keep using
[16:27] [unclear words: ASR "high Q"] tokens? No, I'm sure. Uh, the, that'd be anyways, I'm getting down a weird path.
[16:32] Other actually practical things is, um, one really cool use I've seen for local models,
[16:37] like models that are small and would run on your laptop is I've seen workflows where
[16:41] individuals will send the data to their local model with a prompt to redact the data.
[16:47] And then when they get the data back, they'll automatically send it to like the frontier labs,
[16:51] the redacted version, right? And so there's a couple of tools that will automate that pipeline.
[16:55] One of them is called fabric, um, where you can actually set it up, say like, hey, anytime you
[17:00] see something that looks like it's sensitive, send it to the local model, redact it. But I'm
[17:05] sure there's other tools and workflows you could use. That's just a really great use of a small
[17:08] local model. Um, there have been some really good innovations in the last week on small local models.
[17:13] So specifically the communities come out with a, like a retuned version of the [unclear model name: ASR "quen 3.5"] model.
[17:20] But they've made it down to almost 4 billion parameters, which means they can run on about
[17:25] four gigs of memory. Um, so like on almost everyone's laptops and it's, it's, it's doing
[17:30] really well in coding tasks. So you basically ripped out everything on the model, not related
[17:33] to coding. There's also another small model that's an 8 billion parameter model that came out this
[17:38] week. That's really focused on computer use. And it actually, the local model using the computer
[17:43] scored better than [unclear comparison target: ASR "opening eyes Astra"] does on using your computer. We add a 8 billion parameter.
[17:49] So that's that like, um, you know, something anyone can run it. Most people can run on their
[17:53] home laptops. So yeah. And yeah, it was crazy that it outperformed it. I, you know, you never see
[17:58] that. Um, so, and then the last thing that I just want to say is, um, I liked it and you can do this
[18:04] in a lot of different tools. You can set up like a schedule task or a cron job. And so I just have it
[18:08] review the code that I'm working on every night or what I want to do in the future that I'm going
[18:12] to work towards is actually reviewing my server configurations, like the OS and software configurations
[18:17] too. Cause like when it comes to reaching into servers, right? Like there's typically, you know,
[18:23] three main ways that occurs, right? Like one is like, there's a known vulnerability and you just
[18:29] update the software, right? Two is like, there's a vulnerability in your custom software. And that's
[18:35] where these cron jobs to look for vulnerabilities in your software, in your code will help. And then
[18:39] three, the last one is more like there's a misconfiguration, like a weak credential or a
[18:45] bypass of authentication or something like that. And typically those are more like in the services
[18:48] you're configuring rather than in the code. And so, I dunno, I'm trying to like slowly automate
[18:53] those things for myself. So that way, like ChatGPT or Claude, or I'm using Hermes for a lot of
[18:59] this stuff, just as going through and checking all my servers every day and trying to cover those
[19:03] three angles. Cause I feel like, well, you know, it's nice to talk about zero-days and things like that.
[19:08] Like most of the time, the way people get into your systems is just due to like sloppy code or
[19:14] sloppy configurations. So, or you just forgetting it exists and never patching it. So.
[19:19] Yeah, I was going to say unpatched software is a big chunk. Yeah. Very helpful. I was going to say
[19:25] at the end of the paper, it also gave advice for like [unclear term: ASR "MCP service"] specifically, code review
[19:31] and CI/CD pipelines. And kind of made the point that like Claude Code is not a security tool and
[19:36] doesn't replace those. So also make sure that you're doing a scanning of your code, like you
[19:41] mentioned to make sure that it's up to snuff. Yeah. I think we're rapidly in an era where,
[19:46] and I think this isn't, this is not like a major change, but, but like you want to have defense in
[19:50] depth. You don't want to be reliant on just one thing. Right. And so if you can have tools that
[19:54] are good at using AI to help you secure systems, that's great. But then you also want tools that
[19:57] are like more programmatic in nature. Right. And more like traditional tools. So you want some,
[20:03] you want some good mix between those two. Sweet. Well, have you seen anything fun this week that
[20:10] you want to share, Shelby? Yeah. I found a puzzle and I was a little disappointed because it claims
[20:18] to be 4D. I don't know. But the fourth dimension, like that's time. Now I don't know what the fourth
[20:25] dimension is. I don't really understand why 4D is. It did not go by the time turner,
[20:28] but it was really fun. So it's this Harry Potter puzzle and it's like a puzzle that becomes like
[20:33] the backdrop. And then you do a second puzzle that's like foam pieces that sits on top of it as an
[20:37] overlay. Oh. And then it's got these little cutouts and then you pull them out and you put in like
[20:42] 3D models for like Hogwarts and everything. So that was good fun. That's what I've been up to.
[20:47] That's sweet. I don't have anything near that cool. Sorry, Bryce. Oh no. What do you have to tell us?
[20:56] No, you're good. You're good. But I'm going to go watch Godzilla [unclear words: ASR "blow abilities"] tonight. So that
[20:59] should be rewarding. So that's my weekend plans is go see the new Godzilla movie,
[21:03] uh, which I heard was getting really good reviews. But I guess the reviews,
[21:08] which I just told you the last episode, I never listened to reviews, which I'm a liar then.
[21:11] So the, um, they're saying it's a little sillier than the last one. So we'll see. I'll give you,
[21:16] I'll give you my, this is my pre-watch take. And then on Monday, I'll give you my post-watch take.
[21:20] Okay. Uh, if I remember goldfish. So, um, thanks. All right. Well, I think that's a wrap for us this
[21:28] week. If you, uh, want to stay up to date, come on back. We'll be here again next week. All right.
[21:33] See you. Bye.