Read Time: 8 minutes

Brought to you by:

Howdy friends!

For the first time in 190 weeks I just didn’t get to finish the newsletter Thursday night. So you’re getting this about 2 hours later than normal! I hope no systems came to a halt because of it. I did get a text just last week from a friend that said “Man every time your newsletter shows up I’m surprised it’s already Friday!” - I responded, “How do you think I feel?”

Off to Vegas this week for OneCon put on by SentinelOne. Me and Ed are going to be filming our podcast The Low Down there and interviewing some folks who competed in a threat hunting CTF they put on. Should be a good time! If you’re there, say hi and come find the big glass fishbowl recording studio they have us in on their expo floor.

🖊 Something I wrote: Are we going to lose public key cryptography?

🎧 Something I heard: Jhaddix’s bug bounty village talk at DEFCON

🎤 Something I said: Another great episode of The Low Down talking about Citrix, ShinyHunters, and more

Hardware wallet implants went from conference talk to roughly $100M stolen in a week. Someone has been opening genuine Ledger wallets and adding a tiny board that taps the screen, copies your seed phrase as it's shown to you during setup, and phones it home over its own cellular modem. Bitquery ties about 315 drained accounts to devices from CryptoBilis, an authorized Ledger reseller in Malaysia that quietly changed owners in March under an NDA. Joe Grand reported cruder versions to Ledger back in March. Mark Karpelès (the Mt. Gox guy) later found a polished one with a silk-screened part number, which is what a production run looks like. A German buyer says he's now found them in two Nano Xs bought in person at MediaMarkt.

Trezor responded by cutting resellers while saying there's no evidence its devices were hit. Kim Zetter followed up with Grand's August report to Trezor about two implanted Trezors from sketchy third-party sellers that hand you a seed phrase the attacker already knows. MG put it bluntly: Ledger is downplaying and Trezor is lying. His other point is he built a malicious Ledger with no visible changes five years ago, so cracking the case open and eyeballing it proves very little. Buy direct from the manufacturer, and if you recently bought from a reseller, move funds to a fresh wallet bought direct with a new seed. (read more here and here)

More than half of security teams already use AI pentesting across their environment, and fewer than 1% have no plans to try it. What matters most to them comes down to data privacy, safety on live systems, and being able to check the work. Intruder's State of AI Pentesting 2026 report surveyed 500 security and engineering leaders to find out:

Get the report (no email required!)

*Sponsored

Somebody defrosted France’s food supply. Overnight on October 11, attackers remotely cooked the cold cases at a few hundred French supermarkets. Intermarché says 150 to 200 of its stores were hit and Coopérative U close to 100, plus Leclerc, Monoprix and others. Temps reportedly reached 50C and the Paris prosecutor is reporting fires. Intermarché's CEO told Reuters it cost tens of millions of euros in a single day. The common thread is internet-facing Danfoss refrigeration controllers. Kevin Beaumont says there are thousands of them online, all with a built-in admin login with the password keyed to the current date.

Pen Test Partners laid out how little this took. Getting in is trivial, and spoiling the food meant pinning the defrost heaters on across every case, which the store's management box will happily push everywhere at once while muting the alarms. Synacktiv published the login bypass in 2023, Claroty's Team82 flagged it again in 2025, and Danfoss has since removed it. The boxes were still online, often through a refrigeration contractor's own cellular modem the retailer had no record of. The same weekend, Ahold Delhaize (Food Lion, Stop & Shop, Albert Heijn) pulled systems offline for an incident nobody has tied to the fridges. Time to go find the contractor-installed remote access in your buildings this week. (read more here and here)

The AI reverse engineering wave hit all at once this week. REA (Reverse Engineer Anything) is an open-source MCP server that plugs agents like Claude Code and Codex into Ghidra, IDA and Hopper, so they can work from an app's behavior down to the binary. It sat mostly ignored after launching in April. Then it went from under 10k GitHub stars on October 6 to over 100k by the 12th, per Decrypt. People immediately pointed AI agents at games, with GTA V running natively on an iPhone and Fable 2 booting on PC. Publishers are already reacting. Bethesda sent a cease-and-desist over an AI-built Fallout browser game, and Activision pulled footage of a decompiled Modern Warfare 2.

The other headline grabber is ArtCraft, free Rust rebuilds of Photoshop, Premiere and now Word, Excel and PowerPoint, built with Claude. Its developer told Decrypt it's a clean-room rebuild with no REA involved, and one pixel artist called the Photoshop clone unusable. My illustrious podcast cohost, Low Level, digs into how much of the Adobe hype holds up in his video. For security folks, the takeaway is that anything running on a device you hand to users can now be picked apart with a prompt, so copy protection, license checks and secrets baked into a binary just got a lot cheaper to defeat. Also, a repo this hot attracts impostors. REA's own README warns about lookalike copies, so only pull it from the official GitHub, rea.tools or npm. (read more here and here and watch more here)

TWO stories this week where the people you call during a ransomware incident were allegedly working the other side. The FBI arrested Edward Dubrovsky, a 54-year-old Canadian ransomware negotiator (formerly at Cypfer, lately CyberSteward), while he was in Philly for a cyber insurance conference. The complaint is sealed but the charge is extortion conspiracy, and the case moved to the Eastern District of Texas. Krebs reports that's the hub of the ShinyHunters investigation into the FBI jobs portal breach, though the FBI hasn't officially tied him to the group. He also just published a book on handling cyber extortion…

Separately, DOJ charged Zohar Pinhasi, owner of Florida-based MonsterCloud, with wire fraud. MonsterCloud sold itself as the principled alternative to paying, with "proprietary tools" for decryption. Prosecutors say he quietly paid the attackers and marked it up. In one case he allegedly paid about $8,200 to the hackers and billed the client roughly $150,000, and overall he charged clients $19M against $8M in ransoms. If you have a negotiation or recovery firm on retainer, ask exactly how they get your data back, and get the answer in writing with counsel involved. (read more here and here)

Brendan Dolan-Gavitt and three colleagues at XBOW just collected Google's top Chrome bounty, $250K for a full chain that goes from loading a web page to running code on a Windows box. It uses three bugs, each built as a swappable stage. A JavaScript engine bug gets control of memory, an image decoding bug breaks out of the inner sandbox, and a crash-reporting bug gets all the way outside. The awarded chain came together in six days in Vegas during Black Hat week. Agents running OpenAI's publicly available GPT-5.6 Sol did the source code digging, testing and first exploit drafts, for about $40K in compute.

Before anyone panics, all three were fixed in Chrome 152 back in August and September, XBOW says none were ever used against users, and the "billion users" figure is just Chrome's install base. The six days also sat on top of months of earlier work, and the humans still picked targets, checked output and decided how to stitch the pieces together. Their first chain even got disqualified because Google already knew about one of the bugs. Still, a browser chain like this has historically been a months-long job for an elite team, and here it was a $40K compute bill on a model anyone can buy. (read more here and here)

An OpenAI engineer, 35-year-old Saudi national Abdulrahman Alfozan, was arrested October 9 for allegedly bugging an executive conference room at OpenAI's San Francisco office. According to the DOJ, staff found a recorder on September 30 plugged into a USB port on the back of a TV, wrapped in toilet paper, with encrypted contents. Security put a camera on the TV, and per TechCrunch it caught him coming back to look behind it. A second, similar recorder turned up in a bag he left lying around. He's charged with intercepting oral communications, which carries up to five years.

OpenAI's statement says it used its own models to understand the device and trace it to the employee, and that it sees no sign he touched model weights, source code or customer data. The replies immediately clocked the marketing in that first sentence. Toilet paper and a TV's USB port is pretty humble tradecraft. Couldn’t even use Claude to ask how to do your crime against OpenAI??? (read more here and here)

The jailbreak scene had a very fun few weeks that most of us only heard about after the exploit died. johnny (@0xjohnny) says a bug in one of Apple's server-side services gave root and a working semi-jailbreak on current hardware, up to iOS 27.2 beta 3 on an iPhone 18 Pro Max, with no new kernel exploit needed. They kept it private because a public release would have been patched instantly. Apple patched it anyway. opa334, the Dopamine jailbreak developer, called it the craziest bug he's seen in 10 years and said everyone who knew about it effectively got a Security Research Device overnight.

Apple's research iPhones come from a locked-down program you apply for, and opa334's point is that a lot of people suddenly had that level of access and nothing bad happened. Everyone else has nothing to do here. The fix lived on Apple's servers, so it applied to everyone at once. Phones that already had the environment installed keep what's on them, but they can't push anything new. Some of the most interesting iOS attack surface lives in Apple's cloud, and this one got fixed without a single software update. (read more here and here)

1,413 CVEs in one Patch Tuesday. That beats Microsoft's previous record for an entire year (1,261 in 2020), and the Krebs article says we're now past 4,000 for 2026. Windows alone accounts for 1,102 of them. Microsoft is, you’ll never guess it… pointing at LLM-driven bug hunting for the spike. ZDI's Dustin Childs told CyberScoop that AI is shrinking the time between a flaw existing and somebody finding it. Windows Shell went from 3 patched bugs last year to 26 so far this year.

Only one is exploited in the wild, CVE-2026-96951, a Windows Shell security feature bypass, so patch that first. Kind of wild to think about though right? 1400 bugs and 1 actively exploited. Microsoft also quietly extended free Windows 10 security updates to October 2027. (read more here and here)

I literally need a section of my newsletter called Edge Device Critical Zero Days. They just keep coming. Ivanti Connect Secure season never ends. Researchers caught an attacker chaining zero-days on a customer's appliance starting October 3. An auth bypass plus code execution (CVE-2026-104935) got them in with no credentials, then a privilege escalation took them to root, and Ivanti hasn't fixed that second bug yet. From there they grabbed stored credentials, including LDAP domain creds, dug through Active Directory for passwords and started password spraying. Ivanti calls it limited exploitation, CISA added it to KEV the same day, and Volexity tracks the actor as UTA0576, leaning criminal with low confidence.

Two things make this worse than a normal Ivanti week. Volexity says Ivanti's own Integrity Checker Tool misses this chain, and Ivanti says there are no reliable indicators of compromise, so a clean scan tells you very little. Patch per the advisory, hunt with Volexity's indicators anyway, and rotate every credential the appliance stored, since patching doesn't un-steal your LDAP creds. (read more here and here)