AI Coding Tool Leaked 1,000+ Developers' Passwords to the Public Internet
Saturday, November 7, 2026 · 5-minute read
French cooperative bank Crédit Mutuel Arkéa was hit by a cyberattack that enabled fraudulent transactions totalling up to $133 million. The attackers got in through a service provider, not the bank directly — a reminder that your bank's security is only as strong as every company it works with. The bank had to shut down some of its systems to stop the bleeding.
↗ BleepingComputerSecurity researchers found that computers inside the White House network have been running residential proxy software, including one app tied to a Russian founder whose servers were used in the 2016 US election hacking. Anyone paying a few dollars could route their web traffic through the Executive Office of the President's network. The activity was spotted on 26 out of 30 days between October and November, mostly in the evenings.
�� Krebs on SecurityGoogle worked with partners to take down a massive botnet built from Android TV set-top boxes. The manufacturer, Chinese firm Videostrong, intentionally pre-loaded malware that rented users' internet connections out to criminals — without anyone's knowledge. The takedown also cut access to roughly 10 million other infected devices, including healthcare gadgets and karaoke machines.
↗ Krebs on SecurityHackers started attacking a newly disclosed flaw in OpenSSH within hours of it becoming public. The bug, nicknamed SSHole, lets anyone who knows about it log into accounts that have no password set — without needing a password or key of any kind. It affects Linux systems running OpenSSH on a specific setup called the musl C library, most notably Alpine Linux's default configuration. Security researchers confirmed they gained root access on a test machine using the flaw. A fix is available in OpenSSH version 10.6p2.
Attackers are sending fake invitations to events supposedly hosted by OpenAI, SpaceX, and Anthropic — big names employees are likely to click on. The emails lead to convincing login pages designed to steal both your password and the two-factor authentication code in real time. Two different phishing-as-a-service platforms are powering these campaigns, targeting Microsoft and Google accounts.
An attacker with no account and no password can send an oversized password string to log into accounts deliberately configured without usable passwords. It affects OpenSSH on Linux systems using the musl C library with PAM disabled — Alpine Linux's default setup is vulnerable out of the box. Hackers started exploiting this within hours of it going public, and confirmed they reached root-level control on a test system.
Status: Patch available — update to OpenSSH 10.6p2 now.
There's an actively exploited flaw in WordPress — and even installing the patch can leave you vulnerable for a brief window. During the update process, WordPress unpacks the new version into a temporary folder that's reachable from the web. An attacker who hits that window can run code on your server without needing any account at all. CISA has added this to its catalog of known exploited vulnerabilities.
Status: Patch available, but block web access to your wp-content/upgrade folder before updating to close the attack window.
Yes, Windows Paint. A flaw in the app could let an attacker run code on your computer over a network — no action required from you beyond having the vulnerable version installed. Microsoft rates it Important. The fix is available but won't arrive through Windows Update; you have to get it through the Microsoft Store. Some machines are updating automatically; others aren't.
Status: Fix available — open the Microsoft Store and check for a Paint update (look for version 11.2609.401.0 or later).
OpenAI released a new AI model called GPT-6 Cyber, built specifically for security research. It's already found 81 previously unknown high-severity vulnerabilities in Firefox and built working exploits for 948 separate Linux kernel flaws. OpenAI calls it a "Critical" capability under its own safety framework — meaning the company considers it powerful enough to cause serious harm if misused. The model scores 48.5% on ExploitGym, a benchmark that tests whether AI can turn known software weaknesses into real working attacks, up from 32.1% for its predecessor. OpenAI says it's also significantly better at following boundaries and not attacking systems it's not supposed to — staying in-scope 99.93% of the time — but the gap between AI-powered offense and defense is narrowing fast.
Imagine a sticky note on your front door listing your house alarm code, your bank PIN, and your email password. That's roughly what's happened to over a thousand software projects on GitHub — except the sticky note is a file that Claude Code, Anthropic's AI coding assistant, quietly created and developers accidentally made public.
Here's how it happened. When you use Claude Code to run commands on your computer, it sometimes asks for permission first. If you click "Yes, and don't ask again," Claude saves that command permanently in a settings file called .claude/settings.local.json. That sounds harmless — until you realize that some of those commands contain API keys, database passwords, or other credentials typed directly into the command. If a developer then uploads their project to GitHub without removing that file, those secrets go public. Security firm Apiiro scanned GitHub and found 1,122 public copies of this file holding at least one live secret. One company confirmed its key was still active when researchers contacted them this week.
The situation is worse than it sounds for one important reason: deleting the file doesn't fix it. Git keeps a full history of every version of every file. Even if you remove the settings file today, anyone can scroll back through your project's history and find the old version — with the secrets still in it. And if anyone has already forked your repository, that copy lives on too, beyond your control.
If you or anyone at your company uses Claude Code, the advice from researchers is clear: don't delete the file first — rotate first. That means immediately changing every password, token, and API key that could have ever appeared in a command you approved in Claude Code. Treat them all as compromised. Only after you've replaced all the secrets should you then work on cleaning up the file from your git history. Going forward, never type credentials directly into shell commands — store them in a proper secrets manager instead.