← All articles

Linux kernel bugs, GSocket persistence and AI-led attacks

Linux kernel bugs, GSocket persistence and AI-led attacks

Linux kernel bugs are among the most significant findings in this week’s security roundup, but they are not the only threat facing website owners and IT teams. Microsoft’s scan of selected Linux kernel subsystems produced 26,502 raw findings, which led to 59 public upstream reports and 33 assigned CVEs. Eight of those CVEs had working root privilege-escalation exploits, including CVE-2026-64563. Microsoft described that flaw as a one-line state-reset mistake in rhashtable, a shared resizable hash table in the Linux kernel. Because software trusted object pointers returned during a table walk, the error enabled stable local privilege escalation on Linux and affected Android even when SELinux was enforcing, without special kernel configurations or user namespaces.

Microsoft also reported a 356.2% increase in devices with GSocket-related detections between March and August 2026. GSocket is an open-source remote access tool that attackers placed on internet-facing Linux servers after exploiting web applications. It allowed them to keep access even after the original vulnerabilities were fixed. Because GSocket relays traffic through a shared public network, blocking an attacker-owned IP address or domain may not work. Microsoft advised checking for deployment, persistence, privilege escalation and outbound connections, rather than treating application patching as complete remediation. Most detections involved PHP web applications, cPanel hosts and Apache, with education and research the most affected sector, followed by technology and communications, manufacturing, and government.

A Europol-led Operation Endgame disrupted infrastructure used by GootLoader, CastleRAT and StealC, seizing 235 servers and 257 domains. GootLoader has infected at least 25,000 computers globally since 2018 and was used in ransomware attacks against at least 44 victims in the United States, causing about $28.8 million in losses. Authorities identified more than 6 million stolen victim credentials, which criminals sold for later attacks including ransomware. Of the seized infrastructure, 120 servers and 39 domains were linked to CastleRAT.

Browser-based attacks were also prominent. Palo Alto Networks Unit 42 identified Falcon, also known as XG or Xiàng Gē, a command and control framework that runs malicious JavaScript inside browsers. Falcon shares nearly identical URL structures, payload delivery methods and communication protocols with PeckBirdy, and the two are assessed as variants of the same core framework. Attackers compromised visitors to webmail portals and government web applications, then used the browser to sideload malicious Chrome extensions and steal passwords, emails, session cookies and communications. Falcon also delivered an implant called Antino, which used Microsoft OneDrive as a command and control channel. The activity targeted government entities in one Middle Eastern country and four Southeast Asian countries between March and May 2026, including foreign affairs, defense and interior ministries and law enforcement agencies.

CloudSEK described Silent Lynx espionage activity against Central Asian governments and critical infrastructure between April and October 2026 after an exposed server revealed tools, stolen data and an AI agent used to operate intrusions. The operation ran in Russian, and test traffic came from a residential Kazakh connection, leading CloudSEK to assess with medium confidence that the operator had a Kazakhstan nexus. The targets and collection priorities included transport, defense correspondence and resource ministries, suggesting regional intelligence gathering rather than financial theft.

Ransomware group Citrus targeted Windows corporate environments through compromised contractor accounts. During later stages, the attackers used AI-generated PowerShell scripts, batch files and commands, removed tools and output files, cleared command history and deleted event logs. They also used trusted brand names and benign-looking filenames to disguise their tools. The Go-written Sn00py-family ransomware encrypted files and added the .citrus extension while hindering recovery and incident response. Kaspersky said the activity shared several similarities with Toy Ghouls and warned that limited contractor access can become a route to wider compromise when it is not tightly controlled and continuously monitored.

Proofpoint documented BluePhishing, a development of OAuth consent abuse. OAuth is a sign-in method that lets an application receive permission to access an account. In this case, attackers socially engineered Microsoft Entra users into approving a malicious agent blueprint. That one-time consent could create an agent identity in the target tenant, obtain access tokens and operate with the privileges assigned to the agent. Proofpoint said gaps in visibility and governance around AI agent adoption increase the risk.

Other Microsoft changes were defensive. Mailbox Replication Service added OAuth authentication for migrations between Microsoft 365 and on-premises Exchange, including hybrid and non-hybrid organisations. Starting November 10, 2026, Windows added protections for CVE-2026-118737, a critical remote code execution flaw in Remote Procedure Call Runtime with a CVSS score of 8.1. Remote code execution means an attacker can run commands on another computer. Microsoft said the first phase of NTLM signing improvements ensures authenticated sessions establish a cryptographic session key and negotiate message signing, reducing relay attacks. Windows servers are due to require NTLM signing by default from May 2027.

A separate Entra ID incident showed why deleting a secret is not always enough. A complete application credential for the United Kingdom’s Planning Inspectorate was committed in plaintext to GitHub. Although the commit was removed from the branch, a force push orphaned it while the data remained retrievable through GitHub’s raw object content network for 18 months. ControlPlane said the client secret was deleted about 20 minutes after responsible disclosure on August 26, 2026.

Fake software websites delivered more malware. An SEO poisoning campaign impersonated Chocolatey and Microsoft Activation Scripts to distribute Armorsteal, a Python-based information stealer. Its installer added a Python 3.11 environment and components for remote access, surveillance and theft. The malware could log keystrokes, monitor the clipboard, capture screenshots and show fake cryptocurrency wallet notifications, while stealing browser data, Discord tokens, cryptocurrency wallets, seed phrases, saved passwords, SSH keys and Steam data.

A Grandoreiro banking malware campaign used localised phishing against users in Argentina, Chile, Colombia, Mexico and Peru. The lures were mainly in Spanish, although some used Portuguese in Spanish-speaking countries.

Also available in: DeutschItalianoShqip