Read Time: 8 minutes

Brought to you by:

Howdy friends!

The headlines continue to scream about cyber, and we continue saying “are you new here?” - Every vuln is AI! Every breach is a sign of the “new era” we’re in. If you’ve been here a while, you know the era isn’t new. Obviously things are changing, but the breach trends are not exploding with new techniques. $6 billion stolen due to a fake job interview social engineering malware onto a machine, not agentic swarms.

Hope you all had a lovely Halloween, I’ve successfully stolen most of my kids peanut M&Ms and put them in the freezer for me over the next few weeks. If you did a cool costume share it with me on Twitter or something, I love Halloween.

🖊 Something I wrote: Called it

🎧 Something I heard: Black October!!11!

🎤 Something I said: How is data security and DLP changing

🔖 Something I read: An update on the 9.8 Linux bug we talked about last week and why we’re not seeing mass exploitation

Railway is in the newsletter for the third time, and this week a Railway employee is the attacker. Monday night Josh Manders, founder of hosting platform Primcloud, posted that a Railway engineer had just tried to breach his infrastructure. Per Josh, an account signed up, pulled credentials out of its own container and used them to request every secret on the platform, hit the cloud metadata endpoint, guessed internal database names, and port-scanned the whole internal network twice. About two dozen scripted commands in 15 minutes, all blocked. The engineer, Jake Runzer, replied under his own name: yup, that was me, no malice, my agents testing signups on similar platforms went off the rails.

Railway CEO Jake Cooper's first reply said they'd have to ban non-manual external testing, which sure made it sound like a work thing. By the next day the official statement called it a personal bug bounty side project, run once on his own time with a personal DeepSeek subscription. Primcloud has no bug bounty and its acceptable use policy bans probing without written consent, which Community Notes pointed out right alongside the CFAA.

"My agent did it" is going to be the excuse of the year, and you own whatever the agent you pointed at someone else's network does. Also notice what it went for first: every secret on a hosting platform, the exact thing Railway spent the last two weeks cleaning up after. (read more here, here and here)

Someone in finance just typed their password into a fake Google login page. You’ll hear about it in an hour, maybe. Until then, the attacker is in their inbox, their files, and every app tied to that account. Material Security stops phishing up front, then locks down the email, files, and accounts behind it with agentic controls and automation that act without waiting on you. One bad click stays one bad click.

See how it works.

*Sponsored

Any text over 256 characters is a valid root password. ozgun at Wakefield found that sshd built against musl without PAM (Alpine's default, OpenWrt, and a lot of Docker images) lets anyone into locked or key-only accounts with any password between 257 and 1,024 characters, because musl's password hashing hands back the same marker those locked accounts use when a password runs too long. On Alpine's official AWS and Azure images the default user is one of those accounts and can go to root without a password. glibc, PAM setups and Dropbear are fine, and nobody's seen it exploited. Update to OpenSSH 10.6p2 or turn off password auth today, which you probably should have done already. (read more)

Last week's open question was how the attacker got Binance's signing authority. Binance's initial report, via Reuters, answers it, and it starts back in June, when a wallet team engineer grabbed a tampered tool from a search ad. Binance replaced the laptop, but the team was now mapped. In September a second engineer who was job hunting interviewed for a fake role, video call with the "CEO" of a company that doesn't exist and all, and downloaded the malicious repo. On October 20 the attacker used that machine plus an internal AI tool to take over two wallet developers' systems, and on the 29th moved 233 million XRP out of five hot wallets in four minutes. The last bad transfer landed nearly four hours later, and the FBI blames North Korea.

Same fake interview playbook as always, and two writeups this week show what that "coding challenge" actually does. One developer documented their own after bailing mid-call: the install step quietly started a background implant, and a hidden VS Code task was set to fire the moment the folder got trusted. Ransom-ISAC traced a near-identical lure to a DPRK-linked RAT. Neither is tied to Binance. Run interview code in a throwaway VM with no credentials on it, turn off install scripts and VS Code's automatic tasks, and maybe the laptop with wallet signing access shouldn't also be the one you job hunt on. (read more here, here and here)

Joseph Cox at 404 Media broke the OnlyFake story in early 2024, made himself a fake California license and British passport with it, and used one to get through KYC at OKX. Now he has the founder on the record. Yuriy Nazarenko, a 28-year-old from outside Odesa, went from hand-Photoshopping documents to a site that generated scan-style IDs and phone-style photos of IDs sitting on a surface, built on a template library of real U.S. scans he bought on Telegram. $99 a month got you 10, and $1,500 got you bulk generation from a spreadsheet. The DOJ puts it at more than 10,000 fake IDs and $1.2 million. Customers used them to open bank and crypto accounts and get unbanned on Binance, and a UN panel says North Korean IT workers used it for passports and licenses.

He got caught the boring way. An FBI agent bought IDs undercover in 2024, and Nazarenko sent Bitcoin to an exchange account in his own name. Romanian border officials picked him up in June 2025, he pleaded guilty, and he got 12 months, time served, against a 15-year max, followed by ICE detention and deportation. The confession mostly comes down to him admitting he hid behind "the client decides how to use it." Copycats are still out there, some shut down by the FBI and others popping up. If your onboarding trusts a picture of an ID, assume it's been beatable for about $15 a pop for over two years. (read more here and here)

Vibe coding makes it easier than ever to build and change applications. But do you know what your AI-generated software connects to, what data it accesses, or where the risks are? NexusAI helps teams move beyond static security reviews with AI-powered threat modeling designed for modern development. (read more)

*Sponsored

Three weeks ago somebody defrosted France's supermarkets. This week it went continental. Overnight Monday, cold cases warmed up in stores across Belgium, Luxembourg, Sweden, Ireland, the Netherlands, Estonia and Finland, and Italy, Germany, Romania and the UK reported hits the next day. One ICA in Sweden posted that its freezers had been set to 50C. Sabcobel, which runs Delhaize's refrigeration in Belgium, says the evaporators got flipped to defrost, same trick as France, and that properly protecting the units takes physical modifications on top of software updates. Danfoss told Dutch radio at least 17 companies there were hit. Then France's blood service, EFS, said attackers cranked the temperature and defrost thresholds on three contractor-run cold stores holding plasma, and alarms let techs move it before anything was compromised.

Nobody has claimed it. The EU says these are not isolated incidents, the Pentagon is still looking at the freezer malfunctions on US bases from August, and Sophos' Rafe Pilling puts the odds on state signaling or a lone wolf. Sabcobel's CEO says most of its customers moved their fridges behind a secure connection after France. And Zack Korman spotted one Swedish victim saying it had planned to patch on Wednesday just a few days after they were hacked... Same advice as last time: go find the contractor-installed remote access in your buildings, and get anything that controls temperature off the open internet. (read more here, here and here)

CrowdStrike says China-nexus Vapor Panda pointed ARTEX, an open-source multi-agent pentest framework, at more than 200 organizations in 14 countries starting October 17, all from a single controller over 11 days. Nine confirmed compromises, including a Middle East interior ministry and telecoms in the US, Belgium and the Philippines, plus password spraying against a dozen more. The agents ran mostly on DeepSeek with a second commercial model CrowdStrike couldn't identify, and a human operator stepped in with instructions in Chinese at the decision points. They were also told to skip any box running Falcon, which CrowdStrike was happy to include. DIVD now says ARTEX was behind its September breach too, initial access to root in 12 minutes.

But wait, theres more! Graham Helton flagged and Gabe took apart a backdoor in the ARTEX a lot of people have been downloading. The original repo was deleted in October, and the most popular of its nearly 1,000 forks pulls a prebuilt image from a personal Docker Hub account created two minutes before the fork, almost 8,000 pulls. Hidden in the login is a fake captcha check that hands out unauthenticated code execution as root. Gabe's read is that a chunk of the "distributed ARTEX attack network" is the operators getting owned themselves.

Another bonus, my good homie Jason Haddix shipped SpikeStrip, decoys and traps that waste AI pentest agents' time, with an ARTEX mode. If your red team grabbed ARTEX to see what the fuss is about, check which fork and which image to make sure you didn’t grab the backdoor. (read more here, here and here)

Back in July we covered LG's smart TV app store being riddled with residential proxy SDKs. Here's where those SDKs end up. Layer3 Intel spotted residential proxy nodes exiting through the network block reserved for the Executive Office of the President 236 times between October 5 and November 3, from six IPs, mostly between 5 p.m. and midnight. 28 different proxy services could route through it. Synthient's Ben Brundage tied the traffic to Infatica, whose founder also runs Russian host King Servers, which hosted six of the eight IPs Russian intelligence used against Arizona and Illinois election databases in 2016.

How it lands on a White House network is probably boring. App developers bake proxy SDKs into free mobile and desktop apps for small payments per install, staff bring those phones and laptops to work, and anyone with a cheap proxy subscription and a reseller that skips KYC comes out the other side looking like 1600 Pennsylvania Ave. Spur has found residential proxies on nearly 300 government networks, and Infoblox says 65% of its customers query proxy domains. Block known proxy SDK domains at DNS, alert when your own address space shows up in proxy feeds, and keep personal devices off any network where the source IP means something. The White House didn't respond to Krebs. (read more)

OpenAI built a pipeline that scans the Linux kernel, reproduces what it finds, writes an exploit and drafts a patch, with a human reviewing at the end. One run in mid-October with an unreleased cyber model turned 2,308 candidate findings into 1,724 reproduced bugs and 948 unique working local privilege escalations. Those ran against a stock desktop distro with the standard kernel protections on. The bill was about $1.2 million in API credits, roughly $1,279 per root exploit. The same harness on GPT-5.6 Sol in July found 166, so that's more than 5x in three months on code that got hardened in between. To make it extra weird, 20 of 200 early "successes" got root through a different 0-day than the one they were assigned.

The disclosure side is where this gets real. OpenAI says it has sent hundreds of reports and patches for the critical ones, with the rest going out through Patch the Planet partners. Within 10 days, 61 were already patched upstream, 43 had staging patches and 299 were being discussed in public. It also says the agents came up with 10 exploitation techniques that outside kernel experts hadn't seen, and it's keeping those details private. Their own takeaway is the one to tape to your monitor: for a well-resourced attacker, a low-privileged account on a box that's behind on kernel patches is functionally root. So keep kernel updates on a real cadence, and treat sandboxes and isolation like gVisor and Firecracker as the boundary that matters. (read more)

Community was foundational in launching and propelling my career. Community is the only reason I can stand being in Texas during the summer months. Community is the point. Today, I invite you to embrace discomfort on the road to a more vulnerable you.

Stay safe, Matt Johansen
@mattjay