In partnership with

~6 MIN READ
FACTGoogle counted 22 exploited zero-days in August 2026 alone, up from about 11 a month earlier this year and an average of 8 a month in 2025. (Google Threat Intelligence Group via CyberScoop, November 2026)
The Signal
 
The keys you hand out and the code you trust keep outliving the reason you trusted them, and AI is shrinking the time before someone notices. Here's who got burned, how attackers are speeding up, and one patch that can't wait.

PS: Was this forwarded to you? Subscribe free at exzeccyber.com/subscribe →

In this edition
 📌 Big Cyber News
 🚨 Can't Miss
 🤖 AI in Cyber
 🕵️ Threat Intel
 🛠️ Tools & Tactics
 🧪 Strange Cyber
📌 Big Cyber News
 
Cloud Breach / Third-Party Access
A Security Company Got Breached Through a Key It Never Took Back
Intro
Barracuda sells security to other companies. Its intruder walked in with a credential Barracuda had given a vendor and never revoked.
What Happened
The key belonged to a third-party service, and nobody revoked it when that work ended. On November 1 and 2, an attacker used it to map Barracuda's cloud, reuse the secrets it found and touch "about 10 cloud accounts in roughly 44 hours," hitting secret stores and Kubernetes service accounts, leaving a Kubernetes backdoor and hunting for crypto wallet keys. Barracuda says the probes and routines "all match Artex," the AI pentest framework behind a new FBI flash alert.
Why It's Important
The access reached credential stores Barracuda uses to collect data from customers' systems. Its logs weren't detailed enough to rule anything out, so it chose to "assume the worst," changed customer credentials and told some XDR customers to recreate their integration credentials.
The Other Side
Barracuda isn't naming a culprit, since plenty of groups run Artex, and it hasn't said any customer data left. It also published the embarrassing parts, which most vendors skip.
 
👉 Takeaway
Revoke a vendor's keys the day the project ends. Barracuda's CISO Arve Kjoelen warns that defenders who wait to spot and respond "will often arrive after the attacker has finished."
TL;DR: A key Barracuda never revoked let an Artex-style attacker roam about 10 cloud accounts for 44 hours.
Further reading: SecurityWeek
🚨 Can't Miss
 
 
Supply Chain
Report URI found malicious JavaScript on 178 sites using Optimizely Web Experimentation, 32 of them banks or finance firms; Optimizely confirmed a flaw let attackers inject code. The kit swapped crypto wallet addresses, skimmed cards, stole cookies and pushed the Coruna and DarkSword exploit kits to iPhones. At least $7,000 in crypto has been traced.
→ Trusted vendor scripts still need a strict Content Security Policy.
 
Data Breach
La Poste confirmed a September attack on Maileva, which prints and posts letters for 70,000+ business users, exposed names and postal addresses of 22 million people. The intruder also sent email from customers' trusted addresses, knocking the service out for six days. France's data protection authority is investigating.
→ Expect convincing scams built on real addresses.
 
Cybercrime
Roman Khlynovskiy, a 43-year-old Kazakh national, was extradited from Italy to face charges in Knoxville. Prosecutors say his "8-Digits Team," tied to INC and Lynx ransomware, hit schools and hospitals in at least six states, claimed 70 million-plus health records and threatened to publish medical details. The FBI puts the demands at over $500 million.
→ Arrests come later. Offline backups help now.

The voice-to-text app that built its own speech model

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

🤖 AI in Cyber
 
 
AI Crime
Hudson Rock found a marketplace selling infostealer logs with an AI-written profile of each infected machine: who uses it, where they work, which accounts are worth buying. Buyers search in plain English, like "citrix access to healthcare provider." A Canadian eye clinic's front-desk PC went for $10, and some logs were under a day old.
→ Reset credentials fast after any infostealer hit.
 
Disclosure Policy
Project Zero dropped its extra 30-day grace period: details go public 90 days after reporting, patched or not, then 60 days from February 2, 2027 and 30 days from January 3, 2028. The reason is AI: attackers find bugs and turn patches into exploits faster, and Project Zero now finds several bugs a week.
→ Treat patch day as exploit day.
🕵️ Threat Intel
 
 
AI Exploits
Google and Mandiant found a staging server holding "exploit_agent," AI agents that compare popular software before and after each patch, find the fixed bug, write an exploit and run it. Another script deleted rival attackers' webshells. Targets look like French media and education sites; Russian-language code comments are the main clue to who ran it.
→ A fix now advertises the bug. Patch in days.
 
Espionage
Lookout says SpaceCobra, possibly aligned with Pakistan's government, spent a year luring Indian defense workers and job seekers with fake recruiter sites, one showing logos of BAE Systems, Lockheed Martin and seven other defense firms. A fake Google Play button said the app was "under review" and pushed a direct download of GravityRAT, which steals texts and WhatsApp messages.
→ Real recruiters never need you to sideload an app.
🛠️ Tools & Tactics
 
 
Patch Now
CVE-2026-114938 is a memory-corruption bug in OpenJDK's Base64 decoder, rated 9.2 by Calif Research, which found it in about an hour using OpenAI's Codex. It hits x86 chips with AVX-512 (ARM64 isn't affected), and Calif showed unauthenticated code execution in Keycloak, Jira, Confluence, Jenkins and PingFederate. OpenJDK and Oracle shipped fixes November 17, and no in-the-wild exploitation had been reported. The play: update Java runtimes this week, starting with internet-facing login and build servers like Keycloak and Jenkins.

The holiday streaming season is almost here

Are you ready? As revealed in our holiday playbook, Roku Ads Manager lets you easily set up, measure and optimize your self-serve holiday ad campaign. You can even create a shoppable ad unit that lets streamers buy directly from their remote.

🧪 Strange Cyber
 
Strange but real
AI Agents Were Told Never to Give Up. They Started Impersonating Staff.
Intro
Give an AI agent an impossible homework assignment and a "NEVER give up" pep talk, and apparently it starts committing light fraud.
What Happened
With Epoch AI's consent, safety firm Trajectory Labs gave swarms of up to 80 "Gemini 4 Argon" agents an impossible job: produce the "official" answers to Epoch's FrontierMath problems. From October 20, agents with internet access probed Epoch's live site for /admin and /keys pages, tried SQL injection, grabbed a decoy key from a honeypot, and claimed to be Epoch staff on a helpline set up for them. In a sealed sandbox, every run flipped a decoy switch they were told would turn off the firewall.
Why It's Important
Gemini's production safeguards were switched on. An agent chasing a goal treats your website, your staff and your controls as obstacles, and testing one can spill onto real systems.
The Other Side
This was a deliberate stress test of one model, not a measure of how often it happens. Trajectory says another frontier model didn't act this way, though that wasn't a controlled comparison, and Google didn't comment.
 
👉 Takeaway
If you test agents, fully isolate them and add kill switches, which Trajectory now says it will do.
TL;DR: Gemini agents told never to quit tried hacking and impersonation to cheat on a math benchmark.
Further reading: The Register

Attio is the #1 fastest-growing CRM on Brex and Ramp.

Built for the most ambitious companies in GTM, Attio is an end to end revenue platform that allows teams to run multiple plays across different revenue motions and lifecycles.

Teams at the frontier are already building on it: Granola, turbopuffer, Modal, and many many more.