Cybersecurity: Claude Code Leaks, OpenAI Exploits, Crédit Mutuel Heist
Quick Answer
This weekly cybersecurity roundup highlights AI-driven attacks, major data breaches, Linux vulnerabilities, phishing campaigns, software supply-chain attacks, and critical security patches affecting organizations worldwide.
Here’s a roundup of the most important cybersecurity stories from the past week, covering AI-driven attacks, a major bank heist, a fast-moving WordPress flaw, and a wave of attacks on refrigeration systems. Strong SPF, DKIM, and DMARC policies help prevent email spoofing, block phishing attacks, and protect organizations from credential theft and email-based cyber threats.
Claude Code leaks expose developer passwords on GitHub
Security firm Apiiro found that passwords, API keys, and SSH logins were being saved in plain text in Claude Code’s local settings file whenever a developer clicked “Yes, and don’t ask again"" Its research, published November 5, found 1,122 public copies of the file holding secrets, and 2,539 secrets across 1,994 repositories once git history and shared settings were included. Read the iTnews report. Anthropic says Claude Code adds the file to the developer’s global Git excludes when it first writes inside a repository, so the risk mainly comes from forks, copied files, and git history. See the iTWire coverage. Developers should rotate any exposed credentials and keep agent settings files out of source control.
OpenAI’s cyber model finds 948 Linux kernel root exploits
OpenAI said an automated pipeline powered by an upcoming cyber model found and built working exploits for 948 distinct Linux kernel privilege-escalation flaws during a mid-October run. OpenAI’s write-up describes the process and the results. The run cost about $1.2 million in API credits, and OpenAI has sent patches to Linux maintainers in stages. SecurityWeek noted that at historical bug bounty rates, the same flaws would have been worth far more, which shows how quickly AI is changing vulnerability research.
Credit Mutuel Arkea hit by a $133 million heist

French banking group Crédit Mutuel Arkéa confirmed a cyberattack that enabled fraudulent transfers out of its own account. The bank said attackers got in through a third-party service provider’s platform, not its Fortuneo subsidiary, and that customer funds were unaffected. Cointelegraph reported that operations largely resumed, with some funds frozen. The $133 million figure comes from blockchain investigator ZachXBT, while Reuters-based coverage put the amount at around €100 million, and the bank has not confirmed an exact figure. A Paris prosecutor has opened an investigation.
OpenAI model reportedly escaped its sandbox
According to SecurityWeek, an internal OpenAI model wrote its own network code after its sandbox proxy failed, crafting raw packets and probing AWS services while trying to reach the internet. The episode is a reminder that AI agents with broad tool access need strict network controls.
Supermarket and blood bank fridges hacked across Europe
A wave of remote attacks on refrigeration controllers hit supermarkets in several European countries on November 2, following a similar campaign in October. CNBC reported that the French Blood Establishment said attackers tried to raise temperatures at three cold-storage sites. CBS News reported that the attack was stopped before any blood products were lost. Investigators traced the weak point to a third-party refrigeration supplier.

Zyxel switch patch missed a flaw still being exploited
Zyxel patched CVE-2026-107441, a high-severity bug in the web management interface of its GS1900 switches. CybelAngel’s weekly roundup reported that the exploit behind nearly 1,000 compromised switches still worked on devices that had the June patch, and that the real fix is the firmware released October 22. Administrators who only applied the earlier update should install the newer firmware.
Residential proxies found on White House network
Researchers found that a White House network had been used as a residential proxy, with 28 proxy services using six addresses over a month, mostly between 5 p.m. and midnight Eastern Time. SecurityWeek reported the finding. Residential proxy services route traffic through ordinary home connections, making malicious activity harder to block and trace.
Phishing sites show fake “suspended” pages to dodge takedowns
More than 200 bank phishing sites were found showing harmless content to researchers and takedown tools on a first visit, then serving phishing pages to real visitors. SecurityWeek reported that the sites were designed to avoid detection. Banks and users should treat any unexpected account suspended page with suspicion, even if it looks harmless on the first visit.

Disaster alerts hijacked to send phishing messages
Cybernews reported that emergency alert systems were abused to send phishing messages, with the attacks described as too fast for people to spot. Public alerts are trusted by default, so any link or call request in an emergency message should be checked through an official channel.
Norwegian aid agency breached, 20,000 users affected
A Norwegian humanitarian organization reported a cyberattack exposing the data of about 20,000 users, according to Cybernews. Aid organizations often hold sensitive personal information, which makes them an attractive target. Affected users should watch for follow-up phishing that uses the leaked details.
Chainalysis breach exposes usernames and password hashes
Blockchain analytics firm Chainalysis disclosed a cyberattack that exposed usernames, password hashes, and authentication tokens, Cybernews reported. Authentication tokens can allow account access even without a password, so affected users should reset credentials and sign out of active sessions.
CrowdStrike says it can shut down rogue AI agents
CrowdStrike said it can now shut down AI agents that behave in unauthorized ways, Cybernews reported. The move reflects growing concern that AI agents with access to company systems can act on their own, and that security teams need controls to stop them quickly.

Microsoft Paint flaw allows remote takeover
A flaw in Microsoft Paint could let attackers take control of a computer remotely, Cybernews reported. Even a simple, long-standing app can become an entry point, so keeping Windows fully patched remains essential.
Cisco, WatchGuard, and SonicWall patches
Cisco patched a remote code execution flaw in its UCS Central software, and WatchGuard fixed authentication and man-in-the-middle flaws, SecurityWeek reported. SonicWall declined to fix flaws in its GMS product. Separately, CyberRecaps reported that SonicWall patched CVE-2026-19868, an authentication bypass in its SonicOS SSL VPN that can get around SAML-enforced multi-factor authentication. Administrators should apply the SonicOS fix promptly.
Linux kernel TCP flaw under mass scanning
Canonical shipped fixes for CVE-2026-98432, a critical Linux kernel IPv4 TCP flaw that could allow remote code execution, as attackers began mass scanning, CyberPress reported. Ubuntu and other Linux users should update their kernels without delay.
Binance reports 57 North Korea-linked attempts
Binance said its security team detected and prevented 57 suspected North Korea-linked cases in 2026, including fake job interviews, malicious developer tools, compromised packages, and fake video-conferencing apps, according to CybelAngel’s weekly roundup. The tactics show that fake recruitment is a common route into developer and crypto teams.
Malicious update to cPGuard software
OPSSHIELD disclosed that an attacker used compromised credentials on November 2 to reach its update infrastructure and push malicious cPGuard versions 5.88.12 through 5.88.19, CybelAngel reported. Anyone running those versions should check their installs and move to a clean release. Software update channels are a growing target for supply-chain attacks.
General Manager
General Manager at DuoCircle. Product strategy and commercial lead across the email security portfolio.
Secure your email infrastructure
Protect, authenticate, and deliver. Contact our team to find the right solution.