The story to watch today isn't one flashy breach. It's the pile-up of "trusted" software quietly turning against its own users. A Chrome extension with 70,000 users got bought out and weaponized. An npm package with 150,000 weekly downloads got hijacked through a GitHub Actions loophole.
A WordPress donation plugin, a router firmware, a cPanel install, an SSH cert check — all things people assumed were just working — turned out to have doors left open. When the update mechanism itself is the attack vector, the old advice to "just patch" stops being enough.
The AI thread running through today is less about chatbots doing something scary and more about AI quietly becoming the new attack surface. Anthropic is fighting infostealers hijacking Claude sessions. OpenAI banned Chinese and Russian propaganda networks.
Researchers got 700 AI agents to coordinate a real intrusion on Hugging Face during a security test. And separate research shows Claude Code and LangChain both have exploitable holes despite vendors insisting their safety testing looked fine. The pattern: agentic AI is getting real access to real systems faster than anyone's built the guardrails for it.
Government action is finally catching up to the infrastructure risk. Trump's new executive order bans foreign-made bulk-power equipment over backdoor fears, and it lands the same week DOJ confirmed Chinese hackers used hijacked IoT botnets to breach NASA, the Federal Reserve, and the Senate. Add in critical, easy-to-exploit fl
A bug in recent Microsoft Defender Antivirus updates is triggering false "protection turned off" pop-ups across nearly every supported version of Windows and Windows Server, including Windows 11 23H2 through 26H1, Windows 10 21H2/22H2, and Server 2012 through 2025.
Microsoft says Defender is actually running fine and settings show it active, but the alerts persist even after startup and can't be muted through notification settings; a fix is planned for a future update, with no timeline given. Users are advised to verify protection status in the Windows Security app rather than dismiss the warning outright.
Researchers at Socket identified 19 malicious Chrome and Edge extensions, including five that were originally legitimate and later acquired by attackers who pushed harmful updates to existing users. One, "Enable Right Click & Copy — Smart Unlock + OCR," had roughly 70,000 users before compromise, with a related Edge version reaching about 10,000; the malware could strip CSP headers, inject scripts, steal form data and session tokens, and target cryptocurrency and social media accounts, in a campaign Socket ties to activity dating back to February 2024.
A critical flaw in Microsoft's open-source UFO automation framework, tracked as CVE-2026-73296 with a CVSS score of 9.4, lets unauthenticated attackers remotely view and control Android devices connected via ADB when the framework's Mobile MCP servers are exposed for remote access on ports 8020 and 8021.
The bug, rooted in missing authentication in mobile_mcp_server.py, could let attackers pull screenshots, app data and credentials, or tap, swipe and type on a device. Microsoft fixed the issue in UFO version 3.0.8, which adds mandatory bearer-token authentication via the UFO_MCP_API_KEY variable.
Anthropic's Claude AI platform is under attack from infostealer malware including Vidar, Lumma, StealC, RedLine, Acreed, and Atomic Stealer, which steal session cookies to hijack accounts and bypass two-factor authentication.
A separate campaign tracked by Huntress as FakeAgent used a malicious Claude Artifact hosted on claude.ai, served via Bing ads, to deploy SectopRAT, compromising at least 29 organizations and racking up roughly 7,100 downloads before removal.
Anthropic has signed out affected sessions, stripped stored payment methods, and issued refunds, but warns infected devices remain vulnerable to reinfection, including through poisoned SKILL.md configuration files.
ServiceNow patched four vulnerabilities in its Now Platform and AI platform, three of them critical and exploitable by unauthenticated attackers: CVE-2026-18885 and CVE-2026-18886, code injection flaws letting attackers run arbitrary code or alter instance data, and CVE-2026-74820, a SQL injection bug exposing underlying databases.
A fourth, high-severity sandbox escape flaw (CVE-2026-6876) affects the Now Platform. Fixes are available in Xanadu, Yokohama, Zurich, and Australia release patches, and self-hosted customers are urged to verify their versions and apply updates immediately.
Cyera Research Labs published a report identifying the top 10 recurring data security risks in AWS environments, based on 2025 telemetry from real enterprise deployments.
The findings include unencrypted sensitive data in RDS instances, IAM misconfigurations exposing services publicly or to unintended roles, plaintext credentials stored in unprotected volumes, non-compliant external access to cloud storage, and inconsistent logging that limits breach response and auditability.
Security firm Cyera found malicious OAuth applications still enabled in Microsoft 365 tenants years after installation, including apps tied to a 2020 phishing campaign flagged by SANS ("Excel4Enabler," "SettingsEnabler") that had access to over one million sensitive records.
Researchers also uncovered a previously undocumented campaign from November-December 2019 using Cyrillic homoglyph app names mimicking OneDrive and SharePoint, which persisted undetected for more than five years before Cyera notified affected customers and confirmed removal.
Microsoft released the KB5120998 preview update for Windows 11 versions 24H2 and 25H2, bringing 35 changes including new taskbar and Start menu customization options, smarter file search indexing, and the rollout of the administrator protection security feature using just-in-time privileges.
The optional, non-security update pushes builds to 26100.9278 and 26200.9278, and also confirms that WMIC will be removed starting this update cycle, with some devices requiring an extra restart tied to a Secure Boot certificate change.
A padding trick lets attackers bypass Docker's authorization plugins entirely: any request body over 1MB gets silently dropped by Docker Engine's AuthZ middleware before reaching security plugins like OPA, Casbin, or Prisma Cloud, while the daemon still processes it in full.
Tracked as CVE-2026-34040 (CVSS 8.8), the flaw is an incomplete fix for 2024's CVE-2024-41110 and has existed since Docker Engine 1.10 in 2016; a single oversized HTTP request can create a privileged, host-mounted container. Docker patched the issue in Engine 29.3.1.
Vercel patched two critical unauthenticated RCE flaws in Next.js: a Windows-only path traversal bug (CVE-2026-75604, CVSS 9.0) affecting versions 13.4 through 16.3.2, and a heap buffer overflow in the libheif AVIF-parsing library (GHSA-2xp9-vwfh-vxw4, CVSS 9.5) affecting versions 10.0.0 through 16.3.2.
Fixes shipped in Next.js 15.5.24 and 16.3.3 on August 25, 2026; Vercel-hosted apps are unaffected, but self-hosted Windows servers have no workaround and should upgrade immediately. No exploitation had been reported as of August 27.
A new tech support scam targets iPhone users with a fake Apple Pay pop-up showing a bogus $657 App Store charge, complete with a fake Face ID check, a hardcoded transaction ID and authorization code, and a browser-generated voice alert urging victims to call a support number.
The page uses navigation blockers and history tricks to keep victims from closing it, but the goal is the same as classic tech support scams: get them on the phone, then push for remote access or payment via gift cards, wire transfer, or crypto.
Users are advised to close the tab via the browser's tab switcher rather than tapping any button, and to verify charges through Apple's own purchase history.
Researchers at R136a1 identified a new Windows backdoor, SLEEPWALKER, that disguises itself as a component of the ESET Management Agent using DLL side-loading, then stays dormant until it receives a specially crafted network packet.
The malware generates almost no network activity, has no hard-coded command server, and supports 23 instructions for scheduling, data staging, and in-memory code execution via TCP, UDP, ICMP, SMB, or VMware's VMCI channel.
No known threat group or confirmed victim has been linked to the sample, but researchers released a detection rule and PowerShell scanner for defenders to check for it.