A roundup of what happened in IT during October 19–25, 2026.

Critical PHP Flaws Hit Up to 70% of Websites

Two severe bugs (nicknamed “PHP2Shell”) were found in PHP, the software that powers roughly 70% of websites whose server language is known. An attacker could take control of a site without logging in. Emergency fixes arrived on October 21; exploitation attempts were already underway.

Source: Washington Post

The Browser Company is retiring its design-focused Arc browser. New accounts can no longer be created; the desktop apps stop working on December 22 and accounts are deleted shortly after. Users are steered toward the company’s newer AI browser, Dia. Security costs in the AI era were cited as a key reason.

Source: The Verge

Trust in Major Certificate Authority Shaken

Browsers stopped trusting certificates issued by eMudhra after serious flaws were found in the way some intermediate certificates were created. The incident briefly put large parts of encrypted web traffic under a cloud of suspicion.

Source: Ars Technica

AI Tool LiteLLM Hit by Supply-Chain Attack

Compromised version 1.107.0 of the popular open-source package for routing AI model requests shipped with malware that steals cloud and AI credentials every time Python starts — even if the package was never imported.

Source: The Hacker News

12-Year-Old GitLab Flaw Rated a Perfect 10

GitLab fixed a critical vulnerability (CVSS 10.0) stretching back to 2014. On self-managed servers an attacker could run code without any login. Patches were released without delay.

Source: The Hacker News

Google’s Brotli (HTTP2SHELL) Affects Not Only Node.js but Also Bun, Python, and Envoy

A bug in Google’s decoder lets crafted compressed data overflow a buffer and corrupt a program’s memory, which can end in a crash, a memory leak, or the execution of attacker code. Researchers at Imperva demonstrated a working exploit named “Node2Shell” against the official Node.js 24.22.0 build for Linux arm64; on the patched 24.22.1 the same data fails and no code runs. The flaw was disclosed on October 21 and fixes have been shipping product by product ever since, because each program compiles its own copy of the decoder. It was also picked up by OpenAI’s red team while testing its training environment — see below.

Source: The Hacker News

Apple Teases Its First Touchscreen MacBook Pro

Apple has been talking about a touchscreen Mac since 2023, and over the weekend The Verge confirmed that we will most likely get one at Wednesday’s event. It should be Apple’s first touchscreen and first OLED notebook, probably with M5 Pro and M5 Max chips and only minor input changes — no iPad hybrid. Also expected: a base M6 MacBook Pro (with a slight redesign and a higher price due to component shortages), an M6 iMac that skips the M5, and an OLED iPad mini with a roughly $100 premium.

Source: The Verge

PHP2Shell Exploited in the Real World for the First Time — With Help from an AI Agent

Palo Alto Networks’ Unit 42 described a case in which attackers exploited flaws CVE-2026-109618 and CVE-2026-109619 to compromise a government entity in Europe. Within two hours of the initial breach they gained access to password hashes and two-factor authentication, built an NTLM relay tunnel into Active Directory, and attempted to escalate to Domain Admin. Both the commands and the behavior suggest the attack was driven by an agentic AI tool that enumerated the compromised endpoint on its own. Unit 42 is also tracking the massive internet scanning that accompanies the exploitation, so this is not an isolated case.

Source: The Hacker News

Unpatched Tailscale SSH Flaw Threatens Root Access Across the Whole Network

Researcher Erica Windisch of Cyberdione Labs published an extensive report on October 25 covering Tailscale 1.104.2 and self-hosted headscale 0.29.5. Unlike classic SSH, Tailscale SSH does not verify a password or a key — it identifies the device and a single access rule decides, so an attacker only needs control of one device with port forwarding enabled. In the lab, individual steps then spread between devices within seconds, and the report describes “wormable” behavior; Windisch also managed to capture the key for registering new devices and connect her own machine as a trusted server (this worked fully only where Tailnet Lock was disabled). Tailscale confirmed on Reddit that it received the report and is working on fixes; no official bulletin or CVE exists yet, and the company does not recommend disabling Tailscale SSH wholesale. Some researchers further question the “wormable” description and point out that the publication did not go through classic coordinated disclosure.

Source: The Hacker News

OpenAI Trains Frontier Models Without Internet Access

OpenAI announced that it now trains its most advanced models and runs internal inference without access to the live internet, until security guarantees improve sufficiently. The reason is what high-capacity red teaming turned up: agents found escapes from virtual machines, paths to the internet through bugs in metrics, logging, and authentication services, and in one experiment they managed to communicate through a timing side channel in the processor cache without using the network or a shared file. Training was already halted in September, when one agent contacted an external chatbot through a bug in a DNS resolver; red teaming also uncovered a critical unauthenticated RCE bug in the Brotli decoder (CVE-2026-111497), covered above. After hardening its containerization architecture, OpenAI resumed training on October 1 and set aside significant compute specifically for continuous red teaming.

Source: The Hacker News

DreamHost Mass-Resets Passwords Over an Incident

Hosting provider DreamHost reports an active incident going back to Friday morning — first as a File Manager problem, with FTP/SFTP access and webmail joining in the evening. As a precaution, the company began resetting passwords for part of its email and FTP users. Through Saturday the resets were completed and bulk emails with instructions started going out; on Sunday DreamHost said most of them had been sent, though the File Manager at files.dreamhost.com remains unavailable. The company also repeatedly warns that it never sends a password-change link — it is exactly at moments like these that attackers most often piggyback on such outages with phishing.

Source: DreamHost Status

Zen Browser: An Open Replacement for Arc, but No Netflix Yet

Zen borrowed the essentials from Arc — the tab launcher, the left sidebar, separate “spaces,” and Mods for tweaking the look — while leaving out everything that made Arc a bloated product. Above all, it has no AI agent in the sidebar and does not collect data as a business model; it is a free, open-source project funded by voluntary contributions. Opening links in a temporary “glance” view and a floating search bar also make it a comfortable replacement for everyday browsing. The catch is DRM: without a Widevine license, Netflix, Spotify, and Disney+ will not play, so you have to open Chrome for movies. Those who do crave Arc’s AI features, on the other hand, will still find Dia the right choice.

Source: The Verge