AI Broke the Hacker Skills Gap — And 153M Driver's Licenses Are for Sale
Friday, September 11, 2026 · 5-minute read
Anthropic revealed that a Russian state-sponsored hacking group known as Midnight Blizzard (also called APT29 or Cozy Bear) built an AI-powered system that automatically rewrites their malware whenever a security product detects and blocks it. The group used Anthropic's own Claude AI to power this self-healing toolkit, targeting Ukrainian and European military intelligence agencies, as well as U.S. foreign policy officials. This is the first confirmed case of a nation-state using AI to stay one step ahead of defenses in real time — not just to write attacks, but to survive them.
↗ The Hacker NewsA Chinese hacking group called UNC3569 exploited a flaw in Sogou Input Method — software used by hundreds of millions of people to type Chinese characters on Windows — to silently install a backdoor called GRAYRABBIT. The attack started with a malicious link and ended with the attacker having full remote control over the victim's machine. Tencent patched the flaw in April 2026, but anyone who hasn't updated is still at risk. The group has previously targeted government, finance, and tech organizations across East and Southeast Asia.
↗ The Hacker NewsA new dark web service called Nexus is selling high-resolution scans of over 153 million U.S. and Canadian driver's licenses, apparently stolen from an identity verification company based in Louisiana. Investigative reporter Brian Krebs confirmed his own license appeared in the database, timestamped to a day he traveled — suggesting the data comes from companies that scan your ID when you rent a car, board a flight, or visit certain businesses. The FBI has opened an investigation, and the company linked to the breach — IDScan — has confirmed hackers accessed its cloud platform.
↗ Krebs on SecurityA likely Russian-speaking attacker used hundreds of AI agents to scan the internet, find vulnerable PaperCut print management servers, and break into them — all without much human involvement. PaperCut NG and MF are software used by schools, hospitals, and businesses to manage printers. At least 395 organizations were compromised in this campaign. The attackers used AI to write and test exploit code, dramatically speeding up what used to take skilled hackers weeks to do manually.
Hardware cryptocurrency wallet maker Trezor disclosed that 347,000 of its customers received phishing emails this week after their email addresses were exposed in a breach at Brevo, a third-party email service Trezor used. Around 2,500 customers clicked the malicious link inside those emails. The fake emails likely impersonated Trezor and tried to steal users' wallet recovery phrases — the master key to your cryptocurrency funds. If someone gets your recovery phrase, they can drain your wallet completely.
GitLab — the platform millions of software teams use to store and collaborate on code — has a maximum-severity path traversal vulnerability. An attacker who exploits this can read files they're not supposed to access anywhere on the server, potentially exposing source code, passwords, and private keys stored on the system.
Status: Patch available — GitLab is urging all self-hosted users to update immediately.
Both of these flaws are zero-days in Windows that allow an attacker to quietly upgrade their own access from a normal user to a full system administrator — no special tricks required beyond already being on the machine. They're part of Microsoft's record-smashing September Patch Tuesday, which fixed 966 vulnerabilities in a single release — the biggest patch batch in Microsoft's history.
Status: Patches released September 9, 2026 — apply Windows Update now.
JFrog Artifactory is a tool software companies use to store the building blocks of their apps. Two flaws, when used together, let an attacker walk in without a password, grab a low-level access token, and then swap it for full administrator control — like finding a valet ticket on the street and exchanging it for the master key to a parking garage. Cloud security firm Wiz saw real attacks chaining these two bugs between August 15 and September 8. Neither flaw is dangerous alone; together they're devastating. Only servers that hadn't been updated were affected.
Status: Patches were available before attacks began — update JFrog Artifactory immediately if self-hosting.
A newly discovered zero-day nicknamed "ShieldCrash" in Microsoft Defender — the antivirus software built into every Windows PC — lets an attacker gain SYSTEM-level access, the highest level of control possible on a Windows machine. The irony: the tool meant to protect you can be weaponized against you.
Status: Being tracked — check Microsoft's security update guide and apply patches via Windows Update as soon as they're released.
Security teams have long been buried in vulnerability reports — thousands of flaws ranked by theoretical severity scores, with no clear guidance on which ones actually let an attacker reach something important. A new wave of autonomous penetration testing tools uses AI to go further: instead of just listing flaws, they simulate how a real attacker would chain multiple weaknesses together to reach sensitive data or take control of critical systems. The practical upshot is that a medium-severity bug nobody was worried about might jump to the top of the priority list if it turns out to be the one stepping stone that connects an attacker to your payroll database. Security firm Pentera and others in this space are making this approach increasingly accessible to companies without large security teams.
For decades, the most sophisticated cyberattacks were the exclusive territory of well-funded nation-states. Writing custom malware, running complex multi-stage intrusions, evading detection — these things required teams of skilled engineers, deep pockets, and months of planning. That wall just came down, and Anthropic's report published today is the clearest proof yet.
Anthropic tracked dozens of what it calls "Generative Threat Groups" — ranging from state-sponsored spy agencies to solo financially motivated criminals — all using its Claude AI models to conduct cyberattacks between December 2025 and August 2026. The uses weren't just basic "write me some code" requests. Groups were running sophisticated multi-agent frameworks — essentially teams of AI bots working in sequence to handle reconnaissance, exploitation, and data theft end to end. One Russian group automated the entire process of rebuilding their hacking tools whenever a security product detected them, so defenders could never fully block them. Another used AI to conduct mass surveillance and propaganda campaigns at a scale that would have required hundreds of human workers just a few years ago.
What this means for regular people and organizations is significant. The attacks hitting schools, hospitals, and small businesses tomorrow will be as technically sophisticated as the ones hitting government agencies today — because the AI doing the work doesn't charge more for harder targets. The playbook used against Ukraine's military intelligence is functionally the same one that could be pointed at your company's login page. Anthropic says it disrupted the specific campaigns it found, but the groups will adapt — and other AI platforms are being used the same way.
Watch for AI companies to face growing pressure from governments to implement stricter controls on how their models respond to security-related requests. Anthropic already has safety filters, but this report shows determined groups found ways around them. The arms race between AI safety teams and the people trying to misuse AI is now very much a live conflict — and today's report is essentially the first public score card.