> daily_signal(2026_11_06)
The man behind $15 AI fake IDs laid out his operation, OpenAI walled off its exploit-writing model, Meta bought and is killing a top coding tool, and an offline AI aimed at farmers with no signal.
PickBits Daily Signal · Friday, November 6, 2026
// tl;dr
- OnlyFake founder Yuriy Nazarenko has confessed. His site generated fake licenses, passports and Social Security cards for about $15 in Bitcoin; the Justice Department says it cleared more than a million dollars, and the forgeries opened bank accounts and cleared crypto-exchange checks. A photo of an ID was never strong proof; now it is worth close to nothing.
- OpenAI gated an exploit-writing model behind a vetted-access wall. GPT-6 Cyber built working exploits for 948 Linux-kernel bugs in OpenAI's own, unverified test, at roughly $1,279 each; only approved "Daybreak Red" customers can touch it, and OpenAI alone decides who gets on that list.
- Meta bought Conductor and is shutting its cloud on November 20. Teams using it to run many AI coding agents at once have about two weeks to migrate, and OpenAI moved on Pierre Computer the same day. If you own an AI-coding stack, weigh how easily you could leave.
- A crop-disease AI now runs fully offline on one edge device. PlantVLM-Edge keeps 98.6% of its answer quality at about 2 gigabytes and replies in under a second on a Jetson board, giving a grower with no signal a real diagnosis. It is still a research prototype.
I've watched AI identity fraud get worse all year, and this week the man who actually sold the fakes finally talked. Yuriy Nazarenko sold convincing fake licenses and passports out of a site called OnlyFake for fifteen dollars apiece; he pleaded guilty about a year ago, and in a new interview he walked through exactly how it ran. It fits a pattern we've tracked: in August the UK's safety institute caught an agent inventing fake identities on its own, and in September a verification vendor leaked 153 million ID scans. Those were stolen-data stories. This guy was mass-producing brand-new identities from scratch. Offensive AI also went somewhere I didn't expect. All autumn the fear was an open model anyone could run cheaply. This week OpenAI built a far scarier one and kept the only key for itself.
Then there is Meta buying Conductor and shutting its cloud in two weeks. Big players keep snapping up the dev tools teams can't easily leave, and if you built on one, you are on their calendar now, not yours. And the good one, the reason I like ending there: a crop-disease model small enough to run on a single board out in a field, for a grower with nobody to call. It leans on the same kind of commodity AI as that fake-ID site, which I did not expect and keep chewing on.
A man just pleaded guilty to running a $15 fake-ID factory, and the same week a research team fit a crop doctor onto a chip that works with the internet switched off.
1. The $15 AI fake ID has made "show us your ID" prove almost nothing.
Yuriy Nazarenko, who built the OnlyFake site, has pleaded guilty, and the Justice Department says it cleared more than a million dollars.
The step that was supposed to protect you is the photo. Snap a picture of your driver's license, your passport, a Social Security card, and a bank or a crypto exchange treats it as proof that you are who you say. OnlyFake was built to defeat exactly that. For about $15 in Bitcoin it generated a realistic forgery, down to a staged shot of the document lying on a sofa, and it could spit them out hundreds or thousands at a time. In a November 5 interview, founder Yuriy Nazarenko walked through how it worked; he has pleaded guilty and served roughly a year.
The Justice Department puts the site's take at more than a million dollars, and the forgeries did real work: opening bank accounts, clearing crypto-exchange checks, renting safehouses, incorporating shell companies under invented names. We have tracked this all year from the other direction, the breaches, like the 153 million ID scans a vendor leaked in September. Those leaked real identities; this guy made fresh fake ones cheap to produce. I think about that every time a site asks me to upload mine now.
Why this matters: Look, uploading a photo was always a weak check. Now that you can generate a convincing fake in seconds, it is worth close to nothing, and any service still relying on it is barely protecting anything. That's how someone opens an account in your name before you ever hear about it.
Action this week: Freeze your credit while you're not actively applying for anything (usa.gov/credit-freeze), and switch on transaction and login alerts so a new account in your name reaches you first. If you run a product that checks identity with a photo or scan, stop treating that as proof and move to liveness detection or issuer-validated checks before your next onboarding cohort. I have asked enough vendors "how do you know the ID is real" to know the honest answer is usually that they don't.
404media.co: The confessions of an AI fraud kingpin, OnlyFake owner Nazarenko (November 5, 2026)
404media.co: An AI fraud kingpin's confessions (November 2026)
2. OpenAI built an AI that writes working exploits, then decided who may use it.
GPT-6 Cyber produced working exploits for 948 Linux-kernel bugs in OpenAI's own test, gated to vetted "Daybreak Red" customers.
For months the security worry was about open models, the kind anyone can download and run cheaply. OpenAI just did the opposite. GPT-6 Cyber is built both to break into systems and to propose how to fix them, and rather than ship it freely, OpenAI gated it behind a vetted-customer tier it calls Daybreak Red, reachable only through its API and Codex, with identity and IP checks and instructions to run it in isolated environments under human oversight.
In OpenAI's own evaluation the model reproduced 1,724 Linux-kernel vulnerabilities and built working exploits for 948 of them, against 166 for the prior generation, at roughly $1.2 million of compute, about $1,279 an exploit, and it landed at 48.5% on the ExploitGym benchmark, well above the prior generation. Every one of those numbers is OpenAI's own, with no independent verification, and researchers still disagree over whether models like this find genuinely novel bugs. Treat them as a vendor's self-report, not a measurement. AI writing exploit code isn't new; one company holding the best exploit-writer and the keys to it is.
Why this matters: OpenAI is the only one deciding who gets through that door, and everyone downstream is trusting their vetting holds. I have watched enough "trusted partner" tiers spring a leak to not take that on faith, and a model that writes working exploits is a bad place to be wrong.
Action this week: Ask any provider, in writing, exactly who sits inside their "Daybreak Red"-style tier, what identity and IP controls gate it, and nail down isolated-environment and human-oversight terms before any pilot. And if you defend systems, assume cheap exploit generation against known CVEs is here for whoever holds the key, so patch your Linux-kernel and high-value stacks like working exploits now follow disclosure faster than they used to. I think those numbers are marketing until someone outside OpenAI reproduces them, but the patching holds either way.
the-decoder.com: OpenAI's GPT-6 Cyber can both break into systems and propose how to fix them (November 5, 2026)
the-decoder.com: OpenAI launches GPT-5.6 Cyber to help defenders find vulnerabilities before attackers do (2026)
3. Meta bought the tool developers use to run many AI coding agents at once.
Conductor Cloud shuts down on November 20, giving teams that standardized on it roughly two weeks to move.
If your team runs Conductor to drive several AI coding agents in parallel, the clock that matters is November 20. That is the day Conductor Cloud goes dark, and the desktop app stops getting new models in early 2027. Meta is acquiring the startup, CEO Charlie Holtz confirmed the deal without a price, and anyone who built a workflow on it is now staring at a two-week migration, not a product launch.
The same day, OpenAI said it would buy Pierre Computer, the same move on the same afternoon. We watched Meta restrict its own engineers' use of Claude Code and Codex in July, then pull AI out of performance reviews in September; the models themselves are nearly interchangeable now, so the platforms fight over the tooling instead. You built on someone else's product, and they can retire it whenever the strategy changes, with about two weeks' notice if you happen to be the Conductor team.
Why this matters: A tool you route your whole pipeline through feels permanent, right up until the company that owns it sells and switches off the cloud version on two weeks' notice, which is what the teams on Conductor are living through right now. If the owner walks away, can your team keep running? For Conductor Cloud users, that answer arrives on November 20.
Action this week: Export your Conductor configs now, wherever you plan to land. Then next time you pick a tool like this, favor one with open formats or something you can host yourself, so a buyout is an annoyance and not a fire drill. Every tool that has shut down on me, the migration was never the part that hurt. Finding out the export only existed up until the deadline passed was.
the-decoder.com: Meta buys AI coding startup Conductor as big tech races to lock in developers (November 5, 2026)
techinasia.com: Meta acquires AI coding startup Conductor (November 2026)
Listen to the Daily Signal podcast at pickbits.ai/daily-signal.
4. A crop-disease AI now runs entirely offline on a single edge device.
PlantVLM-Edge gives a grower with no connectivity a diagnosis in under a second, on hardware the size of a developer board.
PlantVLM-Edge is built for the grower with a sick plant, no signal, and no expert for a day's drive. It is a vision-language model squeezed down to run fully offline on one small device, so someone can photograph a plant, ask what is wrong and what to do, and get an answer back, with no cloud connection at any point.
They got it shockingly small. A quantization trick keeps about 98.6% of its accuracy while nearly halving the memory, down to roughly 2 gigabytes, small enough to answer in under a second on an NVIDIA Jetson board. Still a prototype on a dev board, not an app on your phone, and it only covers plant diseases and insect pests. Most crop AI ships your photo to the cloud to get an answer. This keeps the whole model on the device, which is the only place it is any use when there is no signal to begin with.
Why this matters: Every other crop AI I have seen needs a data connection, and a remote farm is exactly where there isn't one. Drop that requirement and it finally works out where the signal doesn't reach, which is where the sick crops nobody can diagnose actually are.
Action this week: Watch for whether this moves from a developer board to field-ready hardware, and whether an extension service or NGO pilots it, because that is the step between a good paper and a tool a grower can hold. If you build edge AI, the recipe here is worth copying: squeeze the model until it runs entirely on a single edge device, because that is what puts a diagnostic in the places with no reliable signal. I will be watching the field test most, because a prototype that impresses on a bench still has to survive a dusty afternoon and a cracked screen.
» What to watch this week
- Whether any bank or exchange actually drops photo-based ID checks now. They have known this was coming for a while. A public guilty plea makes it a lot harder to keep pretending an uploaded photo proves anything.
- Who gets inside OpenAI's "Daybreak Red" tier. And whether anyone outside OpenAI ever reproduces those exploit numbers. Until someone does, I am reading them as a press release.
- How many teams actually get off Conductor before November 20. Two tooling buyouts in one day is the part I would not ignore; the deadline is just the first bill coming due.
- Whether PlantVLM-Edge ever leaves the lab bench. The paper proves it works. The boring part is getting it onto rugged hardware and into a grower's hands, and that is the part nobody rushes to fund.
Tomorrow's signal lands here.