Cyber Security News

1. Google Patches First Actively Exploited Chrome Zero-Day of 2026

Google released an emergency update for Chrome to fix CVE-2026-2441, a use-after-free vulnerability in the browser's CSS font feature handling. The flaw, reported by researcher Shaheen Fazim on February 11, was already being exploited in the wild before the patch dropped. Chrome holds roughly 65% of the global desktop browser market, making the attack surface enormous.

Why it matters: This is the first Chrome zero-day of 2026; Google patched eight in all of 2025. The speed of exploitation underscores why automatic browser updates are non-negotiable.

2. Microsoft February 2026 Patch Tuesday Fixes Six Zero-Days

Microsoft's February Patch Tuesday addressed 58 vulnerabilities, including six actively exploited zero-days: CVE-2026-21510 (Windows Shell SmartScreen bypass, CVSS 8.8), CVE-2026-21513 (MSHTML bypass), CVE-2026-21514 (Word bypass), CVE-2026-21533 (RDS privilege escalation, CVSS 7.8), CVE-2026-21519 (Desktop Window Manager EoP), and CVE-2026-21525 (Windows Notepad). All were confirmed exploited before patches were available.

Why it matters: Six simultaneous zero-days in a single Patch Tuesday ties 2025's record. Organizations that delay patching Windows are now exposed on multiple fronts simultaneously.

3. APT28 Weaponizes CVE-2026-21509 Within 24 Hours of Disclosure

Russia-linked APT28 (Fancy Bear) launched a multi-stage espionage campaign targeting European military and government entities, specifically maritime and transport organizations in Poland, Slovenia, Turkey, Greece, the UAE, and Ukraine. The attackers weaponized CVE-2026-21509, a Microsoft Office vulnerability, within 24 hours of its public disclosure, deploying CovenantGrunt and NotDoor payloads through cloud-based C2 infrastructure.

Why it matters: A 24-hour exploitation window means traditional patch cycles are too slow. Defenders need compensating controls (email filtering, EDR, network segmentation) deployed before patches arrive.

4. BeyondTrust Critical RCE Flaw Exploited to Deploy VShell and SparkRAT

Palo Alto Networks Unit 42 reported active exploitation of CVE-2026-1731, a pre-authentication RCE vulnerability (CVSS 9.9) in BeyondTrust's Remote Support and Privileged Remote Access products. Attackers deployed webshells, VShell, and SparkRAT backdoors across financial services, legal, healthcare, and higher education sectors in the U.S., France, Germany, Australia, and Canada. Over 16,400 exposed instances were identified.

Why it matters: Identity and access management platforms are high-value targets. A CVSS 9.9 pre-auth RCE in a remote access tool is as bad as it gets; patch immediately or take the appliance offline.

5. New APT Group Breaches Government Networks Across 37 Countries

A newly identified cyberespionage group tracked as TGR-STA-1030 (UNC6619) compromised 70 government and critical infrastructure organizations across 37 countries over the past year. The group, reportedly based in Asia, deployed phishing, exploitation kits, custom malware, Linux rootkits, web shells, and tunneling tools against law enforcement, border control, parliament infrastructure, and national telecoms.

Why it matters: The scale (37 countries, 70 victims) puts this on par with the SolarWinds campaign in scope. The toolset's breadth allegedly suggests a well-resourced state sponsor.

6. Sandworm Targets Polish Energy Infrastructure with DynoWiper

Russia-linked Sandworm attacked at least 30 energy facilities in Poland, including wind and solar power plants, by exploiting vulnerable FortiGate configurations. The group deployed DynoWiper to directly damage RTUs, IEDs, and serial devices, targeting both OT and IT infrastructure in a sabotage operation conducted in late 2025 that came to light this week.

Why it matters: This is deliberate infrastructure sabotage, not espionage. The targeting of renewable energy facilities in a NATO member state signals an expansion of Sandworm's operational scope.

7. Lazarus Group Hijacks Developer Supply Chains via npm, PyPI, and GitHub

North Korea's Lazarus Group launched a social engineering campaign impersonating job recruiters to trick software developers into downloading malicious packages from npm, PyPI, and GitHub. The group also weaponized VSCode's auto-execution features and used Polygon NFT contracts as dead drops to replace blocked Pastebin C2 infrastructure.

Why it matters: Developer supply chain attacks scale massively; one poisoned package can reach thousands of downstream applications. The use of blockchain-based C2 makes takedowns nearly impossible.

8. Roundcube Webmail Flaws Added to CISA KEV After Active Exploitation

CISA added two Roundcube vulnerabilities to the KEV catalog: CVE-2025-49113 (CVSS 9.9, deserialization RCE) and CVE-2025-68461 (CVSS 7.2, XSS via SVG). The deserialization flaw had been hidden in the codebase for over 10 years and was weaponized within 48 hours of disclosure. Roundcube has previously been targeted by APT28 and Winter Vivern.

Why it matters: Roundcube is widely used by government agencies and NGOs. A 10-year-old, trivially exploitable RCE is a reminder that legacy code is a ticking bomb.

9. Chinese APT Exploited Dell RecoverPoint Zero-Day Since 2024

Google TAG revealed that a China-linked threat group exploited a zero-day in Dell RecoverPoint for Virtual Machines (CVE-2026-22769) for nearly two years before discovery. CISA added the flaw, a hard-coded credentials vulnerability, to the KEV catalog on February 18. The campaign targeted organizations using Dell's backup infrastructure.

Why it matters: Two years of undetected access via backup infrastructure could mean data exfiltration at scale. Backup and recovery systems are often overlooked in security audits.

10. HoneyMyte APT Enhances CoolClient Backdoor with New Capabilities

Kaspersky uncovered upgraded campaigns from HoneyMyte APT deploying an enhanced CoolClient backdoor with clipboard monitoring, active window tracking, and HTTP proxy credential extraction. The campaigns targeted organizations in Myanmar, Mongolia, Malaysia, Thailand, and Russia with an extensible plugin architecture.

Why it matters: Clipboard monitoring plus proxy credential theft is a potent combo for lateral movement and data exfiltration. Southeast Asian organizations are particularly at risk.

11. Konni APT Deploys AI-Generated PowerShell Backdoor

North Korea-linked Konni used AI-generated PowerShell code combined with evasion techniques exploiting Google and NAVER ad redirection structures. The group's campaigns employed ZIP+LNK execution chains, AutoIt-disguised PDFs, and compromised WordPress servers, targeting developers and blockchain infrastructure across multiple countries.

Why it matters: AI-generated malware lowers the barrier for APTs to produce polymorphic, detection-evading payloads. This is no longer theoretical; it's in active use.

12. WEF Global Cybersecurity Outlook 2026 Warns of Widening Cyber Equity Gap

The World Economic Forum's Global Cybersecurity Outlook 2026 found that 87% of leaders see AI-related vulnerabilities as the fastest-growing cyber risk, while the gap between cyber-capable and cyber-vulnerable nations is widening. Only 13% of Latin American respondents believe their countries can protect critical infrastructure, compared to 84% in the Middle East and North Africa.

Why it matters: Cyber inequity creates systemic risk; attackers target the weakest link. As AI supercharges both offense and defense, organizations without resources fall further behind.

AI News

1. Google Ships Gemini 3.1 Pro with Doubled Reasoning Performance

Google released Gemini 3.1 Pro, achieving a verified 77.1% on the ARC-AGI-2 benchmark; more than double its predecessor Gemini 3 Pro. The model is available across Google's consumer and developer products including Vertex AI, Google AI Studio, Gemini CLI, NotebookLM, and Android Studio.

Why it matters: ARC-AGI-2 measures novel reasoning, the kind that matters for real-world problem solving. Doubling the score in one generation is a step-change, not an increment.

2. Anthropic Releases Claude Sonnet 4.6 with 1M Token Context

Anthropic launched Sonnet 4.6 as the default model for Free and Pro users, featuring a 1 million token context window and new records on SWE-Bench (coding) and OS World (computer use). The release scored 60.4% on ARC-AGI-2, trailing only Opus 4.6, Gemini 3 Deep Think, and a refined GPT-5.2 variant.

Why it matters: A 1M token context window in the default free tier democratizes access to long-context AI. The four-month update cadence shows Anthropic is keeping pace with Google's aggressive release schedule.

3. Google Launches Lyria 3 Music Generation with AI Watermarking

Alongside Gemini 3.1 Pro, Google released Lyria 3, a music generation model capable of producing multi-genre compositions with vocals in multiple languages. The model embeds an imperceptible watermark to identify AI-generated content, available directly within the Gemini app.

Why it matters: Built-in provenance watermarking sets a standard the industry should follow. As AI-generated music floods platforms, attribution infrastructure becomes essential.

4. Alibaba Unveils Qwen 3.5 for the "Agentic AI Era"

Alibaba released Qwen 3.5, positioning the model family explicitly for autonomous AI agent workflows. The release comes amid a wave of Chinese AI model launches timed to the Lunar New Year, signaling continued competition between Chinese and Western labs.

Why it matters: "Agentic AI" is the new battleground. Alibaba's framing, and the Qwen ecosystem's open-weight availability, gives developers an alternative to Western closed-source APIs for building autonomous systems.

5. Zhipu Launches GLM-5: "When Models Become Engineers"

Chinese AI lab Zhipu released GLM-5, an AI model positioned for autonomous software engineering tasks. The launch emphasizes end-to-end coding capabilities, from architecture design to debugging, competing directly with Anthropic's computer-use features and OpenAI's Codex line.

Why it matters: The "model as engineer" framing accelerates the timeline for AI-driven software development. Every major lab is now racing to ship autonomous coding agents.

6. Ant Group Releases Trillion-Parameter Ling-2.5-1T and Ring-2.5-1T

Ant Group open-sourced Ling-2.5-1T, a trillion-parameter model supporting native agent interaction with 1M token context and significantly reduced token usage. Alongside it, Ring-2.5-1T, a hybrid linear-architecture thinking model, achieved gold-tier results on IMO 2025 and CMO 2025 math olympiad benchmarks.

Why it matters: Trillion-parameter open-source models with efficient token usage challenge the assumption that frontier capabilities require proprietary infrastructure. Ring-2.5-1T's math results put it in rarefied company.

7. Cohere Launches Tiny Aya Open Multilingual Models

Cohere released the Tiny Aya family of open-weight multilingual models supporting 70+ languages, including regional variants for Africa, South Asia, and Asia-Pacific. At 3.35 billion parameters, the models run on laptops without internet; trained on a modest single cluster of 64 H100 GPUs.

Why it matters: Offline-capable multilingual AI unlocks applications in linguistically diverse regions where connectivity is unreliable. This is AI for the other 5 billion.

8. Altman and Amodei Refuse to Hold Hands at India AI Summit

At India's AI Impact Summit, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei refused to hold hands during a group photo with PM Modi, raising fists instead while other leaders clasped hands. The moment went viral, capping weeks of escalating rivalry including Anthropic's Super Bowl ad mocking OpenAI's ad plans and Altman calling it "clearly dishonest."

Why it matters: The personal animosity between the two labs' founders isn't just theater; it reflects a genuine strategic divergence on safety, monetization, and openness that will shape how AI is deployed globally.

Active Exploitation Watchlist + Notable CVEs

All CVEs added to CISA KEV, confirmed exploited in the wild, or with high-confidence exploitation evidence from the week of February 17–23, 2026.

Prioritized actions this week

1. Patch BeyondTrust immediately: CVE-2026-1731 (CVSS 9.9) is being actively exploited with VShell/SparkRAT deployment across multiple sectors. Over 16,400 instances are exposed.
2. Apply Microsoft February Patch Tuesday: Six zero-days exploited in the wild across Windows Shell, MSHTML, Word, RDS, DWM, and Notepad. No single workaround covers all; the cumulative update is required.
3. Update Chrome and Roundcube: CVE-2026-2441 (Chrome zero-day) and CVE-2025-49113 (Roundcube RCE, CVSS 9.9) are both confirmed exploited. Both are single-click updates that eliminate critical risk.

The Edge

The 24-Hour Window Is Dead

This week crystallized something defenders have suspected for years: the window between vulnerability disclosure and weaponization has collapsed to near-zero. APT28 turned CVE-2026-21509 into a functioning espionage campaign within 24 hours. Roundcube's deserialization flaw was weaponized within 48 hours and had an exploit listed for sale by day three. BeyondTrust's CVSS 9.9 is being used to deploy backdoors across six countries while 16,400 instances sit exposed.

The traditional patch management cycle (scan, prioritize, test, deploy over 30 days) was designed for a world where exploitation took weeks or months. That world is gone. We're now in a regime where disclosure equals exploitation. The question isn't whether you'll be targeted before you patch; it's how many hours you have.

Meanwhile, the AI arms race is quietly reshaping the threat landscape from both sides. Konni is already using AI-generated PowerShell backdoors. The WEF reports that 87% of leaders see AI vulnerabilities as the fastest-growing risk. And yet the same week, Google doubled its reasoning benchmark scores, Anthropic shipped 1M-token context to free users, and Chinese labs released trillion-parameter open-source models. The tools to build better defenses are shipping faster than ever; but so are the tools to build better attacks.

The organizations that survive this inflection point won't be the ones with the most tools. They'll be the ones that close the gap between "patch available" and "patch applied" from days to hours. Everything else is noise.