(/touch-icon-144.png) (/touch-icon.png) (/touch-icon.png) (/touch-icon-144.png) (/touch-icon-144.png) (/opensearch.xml) (Lobsters) (/favicon.ico) (https://lobste.rs/s/cfzhwf/how_are_you_protecting_yourself_against) How are you protecting yourself against the imminent AI dooms zero day? | Lobsters (/assets/application-1101fa63.css) (/assets/system-system-fd030c9d.css) (/assets/tom-select-00e8031d.css) (/assets/TomSelect_remove_button-4d5c34b9.css) (/) (Lobsters (Current traffic: 9%)) (/active) Active (/recent) Recent (/comments) Comments (/search) Search (/login) Login (/login) Login (/login) 10 (/s/cfzhwf/how_are_you_protecting_yourself_against) How are you protecting yourself against the imminent AI dooms zero day? (Above all this is a fun discussion. Don't get too serious :) --- As LLMs get better and better at pattern matching against vulnerabilities, and gaining better logical inference, it's a matter of time until a human or a group of humans and an LLM(s) find a plethora of zero days, some possibly unpatchable. I'm sure others have thought the same, so I'm curious just what the general zeitgeist is. I personally think any machine you haven't put online, like those 10 year old laptops in the...) (/s/cfzhwf/how_are_you_protecting_yourself_against) ☶ (Ask Lobsters) (/t/ask) ask (Developing artificial intelligence, machine learning. Tag AI usage only with `vibecoding`.) (/t/ai) ai authored by (/~LenFalken) (LenFalken avatar) (/~LenFalken) LenFalken (2026-04-21 08:53:02) 5 days ago | (/s/cfzhwf/how_are_you_protecting_yourself_against#comments_story_01kpr51qc8ec2schy3sq8vxzjq) 25 comments Above all this is a fun discussion. Don't get too serious :) As LLMs get better and better at pattern matching against vulnerabilities, and gaining better logical inference, it's a matter of time until a human or a group of humans and an LLM(s) find a plethora of zero days, some possibly unpatchable. I'm sure others have thought the same, so I'm curious just what the general zeitgeist is. I personally think any machine you haven't put online, like those 10 year old laptops in the closest, are safe. The problem is, I believe airgap will not save us either if a truly advanced system is developed. More or less I think "primitive technology" could return in various ways depending what we're talking about, like in finance, to avoid manipulation. (DfDN5vwEPNtLmu1EXlWBKNdHnXLzcgWXOj0auO9CuEWa4C3aP_WURknTsHcMN8reEgugYjqQowIFditvblUjlw) (cfzhwf) (post) (You must be logged in to leave a comment.) (Post) Preview (37) (/login) 37 (/~david_chisnall) (david_chisnall avatar) (/~david_chisnall) david_chisnall (/c/qnepau) (2026-04-21 11:14:32) 5 days ago I’ve written about this before, but TL;DR: New vulnerability-discovery techniques often find bursts of new vulnerabilities. Static analysis, fuzzers, and so on suddenly made it easier to find vulnerabilities. But then you’ve found all of the vulnerabilities that are easy to find with that tool and you reach a new plateau. (28) (/login) 28 (/~dkl) (dkl avatar) (/~dkl) dkl (/c/6ogjjb) (2026-04-21 09:57:03) 5 days ago I think this scenario is actually somewhat far-fetched. Great marketing move, and it will/would be hilarious for the AI charlatans to start extorting the security thespians. Between the two of them there could be a whole pantomime IT industry that literally does nothing. The 10 year old laptops in your closet are only as safe as their most recent update. It's not like ten-year-old vulnerabilities went away. In the (IMO) unlikely event that some kind of security apocalypse happens, I would be delighted if it forces our industry as a whole to turn its attention away from our present obsession with shipping garbage as quickly as possible and towards building stronger foundations and building more deliberately. But that seems pretty unlikely. I can dream though. (19) (/login) 19 (/~dzwdz) (dzwdz avatar) (/~dzwdz) dzwdz (/c/fcoyin) (2026-04-21 10:42:35) 5 days ago I believe airgap will not save us either if a truly advanced system is developed. Unless my airgapped laptop came pre-infected with malware, then just about the only way this could happen is a metallic Arnold Schwarzenegger lookalike coming into my flat and going "Your laptop. Give it to me. Now". Let's be real for a second, this isn't sci-fi. (2) (/login) 2 (/~elephantium) (elephantium avatar) (/~elephantium) elephantium edited (/c/i0218t) (2026-04-21 13:52:08) 5 days ago You might be oversimplifying. After the initial setup, do you truly never need to transfer anything between the airgapped laptop and any other system? Never plug in a USB device? If you have a way of transferring data over the gap, you have a way of inadvertently transferring malware. For that matter, do you mind sharing what you use that airgapped laptop to do? Personally, I have a desktop (daily driver, on the internet) and a laptop (also on the internet, mainly used for job interviews). I'm having trouble coming up with a use case for an airgapped machine. Games? Well, I still need to install them. Watching movies? I need to get those onto the machine somehow (hmm, maybe if you rely exclusively on DVDs). Art? Writing? Well, I still need a way to share my work. I suppose you could rely on printing everything out. It seems awkward, though. (1) (/login) 1 (/~dzwdz) (dzwdz avatar) (/~dzwdz) dzwdz (/c/adoj5o) (2026-04-21 15:10:21) 4 days ago I don't have any "truly" airgapped devices. The closest I got are a few Tails USBs that I never boot up with network access, but I don't think that counts. I just wanted to point out the silliness in the OP. (13) (/login) 13 (/~ndegruchy) (ndegruchy avatar) (/~ndegruchy) ndegruchy (/c/iyp18c) (2026-04-21 09:11:49) 5 days ago Nothing. I'm doing nothing. I already have backups (online and off) and while I'm sure such a scenario would cripple things, I'm not too worried about it. For those of us born before the 90's, we still functioned without Internet, computers, and stuff, it was just different . If AI blows up all the things and we have to go back to ~90's minus the Internet, then I'd say we're doing fine. (7) (/login) 7 (/~eblume) (eblume avatar) (/~eblume) eblume (/c/peikcq) (2026-04-21 11:04:19) 5 days ago Sometimes I feel like it would even be refreshing. Maybe I'd build a mesh network with my neighbors. (8) (/login) 8 (/~madhadron) (madhadron avatar) (/~madhadron) madhadron (/c/mmzzk7) (2026-04-21 11:27:44) 5 days ago it's a matter of time until a human or a group of humans and an LLM(s) find a plethora of zero days That's a very strong assertion and probably not true. Nor is it likely to be a doomsday, since we've already been living in a world where defense in depth is necessary. LLM based phishing concerns me far more. I believe airgap will not save us either if a truly advanced system is developed Why would that be the case? An airgap is breachable now, just like any security measure, but it requires a lot of focused logistics to do so. A source of security vulnerabilities isn't going to drastically reduce the cost of that. (1) (/login) 1 (/~k749gtnc9l3w) (k749gtnc9l3w avatar) (/~k749gtnc9l3w) k749gtnc9l3w (/c/pjjvk6) (2026-04-21 16:14:54) 4 days ago a human or a group of humans and an LLM(s) find a plethora of zero days That's a very strong assertion and probably not true. For some definition of plethora of zero days, and for LLM being Firefox in-browser translator (not used in the process), it is already true now! (7) (/login) 7 (/~cadey) (cadey avatar) (/~cadey) cadey (/c/xexixq) (2026-04-21 11:33:12) 5 days ago I plan to self immolate out in a park so that I am free of this madness forever. (5) (/login) 5 (/~andrewrk) (andrewrk avatar) (/~andrewrk) andrewrk (/c/gkb2qd) (2026-04-21 23:33:39) 4 days ago Don't take the black pill. (2) (/login) 2 (/~h3mulen) (h3mulen avatar) (/~h3mulen) h3mulen (/c/psxf7u) (2026-04-21 12:58:24) 5 days ago Same. (6) (/login) 6 (/~mhd) (mhd avatar) (/~mhd) mhd edited (/c/3ql7hg) (2026-04-21 10:19:09) 5 days ago If having 10 year old laptops around is the solution, I'm not only safe, I'll be the king of the post-CyberdAIne world. Look upon my ThinkPad T42 ye mighty and despair! Spread those photocopied Perl listings, just beware that no one hides malware in a regex! Also, yellow belt in TKD, 30+ years ago. (5) (/login) 5 (/~algernon) (algernon avatar) (/~algernon) algernon (/c/nrtzbl) (2026-04-21 12:17:45) 5 days ago I will employ the tactic of the Lazy Bastard, one that has served me very well for the past few decades: I will do exactly nothing I wouldn't otherwise do, and simply wait for the silly bubble to burst. No bubble, no doom. (6) (/login) 6 (/~david_chisnall) (david_chisnall avatar) (/~david_chisnall) david_chisnall (/c/0m6ozh) (2026-04-21 16:16:29) 4 days ago For anyone who reads press releases from a vendor of bullshit generators and doesn’t assume bullshit, I’d recommend (https://www.flyingpenguin.com/the-boy-that-cried-mythos-verification-is-collapsing-trust-in-anthropic/) reading this article, which describes the particular scent in forensic detail . (3) (/login) 3 (/~andrewrk) (andrewrk avatar) (/~andrewrk) andrewrk (/c/usuiu4) (2026-04-21 23:32:59) 4 days ago This reads like slop. (2) (/login) 2 (/~k749gtnc9l3w) (k749gtnc9l3w avatar) (/~k749gtnc9l3w) k749gtnc9l3w (/c/ggfuyh) (2026-04-22 06:48:51) 4 days ago (https://blog.mozilla.org/en/privacy-security/ai-security-zero-day-vulnerabilities/) https://blog.mozilla.org/en/privacy-security/ai-security-zero-day-vulnerabilities/ — I guess now we have a claim about novel bugs found, signed by Firefox CTO. The claim is 22 security-relevant bugs found via Opus 4.6 fixed in FF 148, and 271 vulnerabilities found via Mythos preview versions fixed in FF 150. No severity classification included in the announcement. About the other points: Mozilla blog post doesn't feel like this writing aims to distinguish «better model» and «better harness». The post claims that the bug types are the same as adversaries have exploited for years, Claude Mythos just, according to the post, found more of those bugs per unit of human effort than normal audits. I do not see any comparison of token/nominal-cost budgets of the two scans. (3) (/login) 3 (/~david_chisnall) (david_chisnall avatar) (/~david_chisnall) david_chisnall (/c/a6kuuu) (2026-04-22 07:05:08) 4 days ago Yup, it was an interesting read. They mention defence in depth and it's not clear how many of the fixed bugs were ones that were in the sandbox and prevented sandbox compromise but were also mitigated via defence in depth. My experience is that people tend to be a bit less careful of security in things they know run with reduced privileges, so that kind of thing can easily sneak in. I'm also surprised at how low the absolute numbers are. 271 sounds like a big number if you forget how absolutely massive web browsers are. I recall a paper around ten years ago looking at FFI bugs between JavaScript and the DOM in Chromium. It referenced around 300 bugs, of one bug class, in one subsystem of Chromium. That's partly why the JavaScript VM boundary isn't treated as a security boundary in modern browsers (that, plus speculative side channels, which mean that the process boundary is the only defensible perimeter). (3) (/login) 3 (/~thesnarky1) (thesnarky1 avatar) (/~thesnarky1) thesnarky1 (/c/p6m8mu) (2026-04-21 18:22:13) 4 days ago Littering my systems with hidden prompt files to make it clear that the SuperMalWare that I am actually an AI in disguise as an advance scout amongst the humans. It highlights that multiple bits of malware running in this house would give up the game and that it should back away quietly. Don't get too serious :) (2) (/login) 2 (/~LenFalken) (LenFalken avatar) (/~LenFalken) LenFalken (/c/k7wxfj) (2026-04-21 20:16:28) 4 days ago Genius (3) (/login) 3 (/~aka_dude) (aka_dude avatar) (/~aka_dude) aka_dude (/c/ol0ton) (2026-04-21 11:11:05) 5 days ago Thinking of using QubesOS on my workstation. Currently writing my own syscall tracer on top of ptrace to semi-interactively allow/deny access to network, filesystem and such. That won't help if the kernel is vulnerable though, which it certainly is. Half-jokingly thinking of using ScultOS locally (probably won't happen). (3) (/login) 3 (/~addison) (addison avatar) (/~addison) addison (/c/uhzxwn) (2026-04-21 11:47:57) 5 days ago Literally nothing has really changed IMO... Why work with an unreliable exploit and then have to conduct secondary attack steps when you could just by creds? The calculus for attackers hasn't changed much, really, and the wildly vulnerable things (e.g., shell injection) were already just waiting to be found. (3) (/login) 3 (/~FlameFlag) (FlameFlag avatar) (/~FlameFlag) FlameFlag (/c/3rmzca) (2026-04-21 17:07:27) 4 days ago Even before AI, I always felt some security (and privacy) folks are a bit too paranoid relative to what the average user actually does My attack surface is pretty narrow: I don't visit many indie sites, most of my browsing is on top-1000 domains, I keep my browser (and all other software) up to date, and I run uBlock Origin. I rarely download arbitrary files outside of images and videos A flood of AI-found zero-days scales with exposure. If the surface isn't there, the multiplier isn't either. So either the software I use is reasonably secure, or I'm using it in a way where security isn't really the bottleneck Sure, it can feel a bit uneasy, and I'll probably be a touch more wary going forward. But I don't want security to become a personality trait of mine, not because I don't care, but because I want the attention budget for other things. Paranoia isn't free I've been fine so far and I expect I'll continue being fine (2) (/login) 2 (/~7tehdt3cnw6kir6o) (7tehdt3cnw6kir6o avatar) (/~7tehdt3cnw6kir6o) 7tehdt3cnw6kir6o (/c/wpnfmt) (2026-04-21 15:17:56) 4 days ago It's not like we've never had to deal with vendors having bad patch cycles, mercenary spyware companies and fuzzers cheapening exploit capabilities, or supply chain attacks popping up in nominally trusted software before. But an airgap isn't really a magic solution. It can have it's uses, but take your system offline. Great. What then? What are you doing with it? How are you getting data in and out? Sure there's data diodes and such but are you sure you're not just reinventing an online system with extra steps? There are countermeasures, but it's mostly an extension of stuff I was doing or planning to do already (ironically, half spurred by the proliferation of vibecoding). I was already using QubesOS after trying and failing to reasonably lock down dev work on windows and linux, moving to remove unneeded apps and services, and sandboxing my homelab in VMs / gvisor containers. I haven't gotten around to building everything on my own CI/CD to mitigate github actions attacks instantly pwning me, or properly microsegmenting my home network, but that's on the todo list. I did recently get buildah to work building containers inside of a gvisor sandboxed container, so that's nice as I don't have to use bare metal runners or pass through the docker socket. It's not a mitigation per se, but I believe we should be donating more to FOSS organizations. They're way better positioned to make meaningful change than me, and underfunded and overworked maintainers lead to things such as the xz attack slipping through. Also if a critical maintainer retires with no replacement there goes any hope of patching anything. Realistically there's only so much I as an individual can do. I cannot harden my bank's infra, nor my dentist's office, nor my health insurance provider, nor my cellular provider. (2) (/login) 2 (/~k749gtnc9l3w) (k749gtnc9l3w avatar) (/~k749gtnc9l3w) k749gtnc9l3w edited (/c/uhjrz5) (2026-04-21 18:27:36) 4 days ago I will ponder the curious fact that the civilisation keeps surviving simultaneous existence of MetaSploit Framework and Shodan. And also keeps surviving NSA that keeps NOBUS 0-days for own use but doesn't always suceed at either preventing or detecting those 0-days being obtained by active adversaries. (/about) About (/tags) Tags (/filters) Filter (/moderations) Moderation Log