Google has released an emergency security update to address a high-severity zero-day vulnerability, identified as CVE-2026-2441. The flaw is confirmed to be actively exploited in the wild, allowing for arbitrary code execution.
Vulnerability Details
Affected Product: Google Chrome versions prior to 145.0.7632.75 (Windows/Mac/Linux)
Identifier: CVE-2026-2441
CVSS Score: 8.8 (High)
Exploitation Status: Actively Exploited
Risk & Impact
Triage: Critical. Immediate patching is required due to active exploitation.
Attack Vector: A remote attacker can exploit this by convincing a user to visit a specially crafted malicious webpage.
Ease of Exploit: The vulnerability can be triggered via a crafted HTML page, suggesting a low complexity for attackers once an exploit is developed.
Action Plan
Immediate Action: Update to Google Chrome version 145.0.7632.75/76 for Windows and macOS, or 144.0.7559.75 for Linux.
Workaround: No specific workarounds have been provided. Users should ensure automatic updates are enabled in Chrome.
Detection: Google has not released specific Indicators of Compromise (IOCs). Enterprises should monitor for and update outdated instances of Chrome.
Relevant professional terms
Zero-day
A vulnerability that is discovered and exploited by attackers before the software vendor is aware of it or has released a patch.
Use-After-Free
A type of memory corruption flaw where a program continues to use a pointer to a memory location after it has been deallocated, potentially leading to arbitrary code execution.
Hobbyist Uses AI to Compromise DJI IoT Infrastructure
Executive Summary
A security researcher has demonstrated the "democratization of exploitation" by using an AI coding assistant (Claude Code) to compromise DJI IoT infrastructure. The researcher identified a critical lack of access controls in DJI’s MQTT message broker, allowing unauthorized access to global device telemetry.
Vulnerability Details
Affected Product: DJI IoT devices (e.g., Romo vacuum, Power stations)
Mechanism: Improper Access Control in MQTT Protocol.
Triage: High. The vulnerability exposes real-time location data and live camera feeds of active devices.
Attack Vector: The researcher used Claude Code to reverse-engineer the mobile app, discovering that the MQTT broker allowed wildcard subscriptions (#) with a valid user token, granting access to the entire fleet's data.
Strategic Significance: This incident proves that AI coding tools can lower the barrier to entry for complex reverse engineering and protocol analysis, effectively arming "hobbyists" with offensive capabilities.
Action Plan
Immediate Action: DJI users should ensure their devices are on isolated guest networks until vendor patches confirm strict topic-level access control enforcement (ACLs) on the MQTT broker.
Mitigation: Implement strict Network Segmentation. IoT devices should never reside on the same VLAN as sensitive corporate data.
Detection: Monitor for unusual network traffic patterns originating from the device, connections to unknown IP addresses, or unexpected device behavior.
Relevant professional terms
MQTT (Message Queuing Telemetry Transport)
A lightweight messaging protocol often used in IoT. Insecure configurations can allow unauthorized users to subscribe to all messages passing through the broker.
Democratization of Exploitation
The trend where advanced offensive capabilities become accessible to low-skill actors via automated tools or AI assistants.
Kaspersky has identified Keenadu, a sophisticated backdoor embedded in Android tablet firmware and spread via Google Play apps. The malware provides attackers with remote control and has exposed operational links between major botnets like Triada and BADBOX.
Key TTPs
Initial Access: Compromising device firmware during the manufacturing supply chain and distribution through malicious apps on Google Play.
Execution: Injects itself into the Zygote process, allowing it to hook into every application launched on the device. Payloads are decrypted and loaded from a core Android library.
Defense Evasion: Modifies core system libraries (libandroid_runtime.so) and establishes persistence within the firmware, making it difficult to remove.
Campaign Analysis
Keenadu represents a significant evolution in Android threats by successfully leveraging a supply chain compromise for pre-installation. Its shared infrastructure and code reveal a collaborative ecosystem among previously distinct major botnet operators.
Targeting & Infrastructure
Target Profile: Users of Android tablets, including Alldocube models, with infections peaking in Russia, Japan, Germany, and Brazil.
Infrastructure: Utilizes shared Command and Control (C2) servers with other botnets, such as the Triada and BADBOX botnets.
A new infostealer campaign targets the OpenClaw AI assistant framework to steal files containing API keys, authentication tokens, and other secrets. This activity highlights a growing trend of threat actors targeting AI development environments to compromise sensitive credentials.
Key TTPs
Initial Access: Infostealers are commonly delivered via phishing emails with malicious attachments, cracked software downloads, or malicious ads.
Execution: Malware often executes commands via scripting interpreters and may use legitimate processes, like a browser running in headless mode, to download additional payloads.
Defense Evasion: Attackers use obfuscated files and may hide malicious code within images (steganography) or inject it into legitimate system processes to avoid detection.
Campaign Analysis
This campaign signifies a tactical shift towards targeting AI-related infrastructure. The theft of API keys and session tokens is particularly damaging, as it can bypass multi-factor authentication and lead to significant financial abuse of cloud services.
Relevant Terms
Infostealer Malware: A type of malicious software designed to covertly gather sensitive information, such as login credentials and financial data, from a victim's computer.
Authentication Tokens: Digital keys that allow a user or application to access resources for a limited time without re-entering credentials, often targeted by malware to hijack active sessions.
New research reveals that Large Language Models (LLMs) are vulnerable to side-channel attacks, where sensitive data is inferred not from the direct output, but from metadata like response timing and hardware memory usage.
Key Findings
Remote timing attacks against models like ChatGPT and Claude can infer conversation topics with over 90% precision by monitoring encrypted network traffic response times.
Hardware cache side-channels can be used to steal both user input and model output by analyzing memory access patterns and the timing of autoregressive decoding.
The Bottom Line
The underlying infrastructure hosting LLMs represents a critical and subtle attack surface. Securing the model's API is no longer sufficient; leaders must now mitigate risks from the physical and virtual environments to prevent data leakage through these indirect channels created by performance optimization techniques.
Relevant Terms
Side-Channel Attack: An attack that exploits information leaked from a system's physical implementation (e.g., power consumption, timing, sound) rather than directly attacking its software.
Large Language Model (LLM): A type of artificial intelligence model trained on vast amounts of text data to understand and generate human-like language.
Researchers discovered a technique where attackers can abuse web-based AI assistants as command-and-control (C2) relays, effectively hiding malicious traffic within legitimate, trusted enterprise communications. This method allows malware to receive instructions from platforms that are typically allowed by corporate security policies.
Key Findings
The "AI-in-the-Middle" technique was demonstrated against AI platforms like Microsoft Copilot and Grok.
Attackers leverage AI assistants with web browsing or URL fetching capabilities to act as a proxy for C2 communications.
This method allows malware to become prompt-driven, dynamically adapting its behavior based on AI model outputs instead of static instructions.
The Bottom Line
This research highlights a critical flaw in the "allow by default" security posture many organizations have for major AI service domains. Security leaders must now treat traffic to trusted AI platforms with the same scrutiny as any other egress traffic. It is no longer sufficient to simply whitelist domains, as they can serve as covert channels for advanced, adaptive malware, challenging traditional network security and detection models.
Relevant Terms
C2 (Command and Control): The infrastructure (servers and software) used by attackers to communicate with and control compromised devices (bots, implants) in a target's network.
Egress Traffic: Network traffic that originates from within a private network and is destined for an external network, such as the public internet.
Polish police have detained a 47-year-old man with suspected ties to the Phobos ransomware-as-a-service (RaaS) operation. The arrest is part of a broader international crackdown on the ransomware group.
TTP 2: Stealing credentials, credit card numbers, and server access data.
TTP 3: Employing double extortion tactics by threatening to publish stolen data.
The Consequence
Outcome: The suspect was detained and faces a potential prison sentence of up to five years.
Assets Seized/Forfeited: A computer and mobile phones believed to be used in the offenses were recovered.
Strategic Takeaway
This arrest highlights the continued success of coordinated international law enforcement operations in dismantling ransomware affiliate networks piece by piece.
Relevant Terms
Ransomware-as-a-Service (RaaS): A subscription-based model that enables affiliates to use pre-developed ransomware tools to execute attacks.
Remote Desktop Protocol (RDP): A protocol that allows a user to connect to another computer over a network connection, which threat actors often exploit if poorly secured.
