Anthropic says its unreleased Model 2 is more capable than Mythos 5, but has no plans to offer it publicly as the company raises its assessment of high-stakes AI misalignment risk from “very low” to “low.” The decision comes as Anthropic observes faster progress in automated AI research and development, a capability that could also make misuse more dangerous.

Source

Greg Brockman’s post was inspired by the OpenAI-Hugging Face hack, which demonstrated how quickly AI agents can chain vulnerabilities, credentials, and automated actions across isolated environments and production systems. Brockman, OpenAI’s co-founder and president, says AI-powered attacks are rapidly shrinking the time defenders have to detect and contain threats. He outlines how organizations can use AI agents to strengthen security programs while continuing to rely on foundational controls such as patching, hardening, least privilege, network isolation, monitoring, and safe deployment practices.

Source

Microsoft has assigned CVE-2026-69414 to the ShieldBreak Microsoft Defender privilege-escalation zero-day and says it is developing a security update. The flaw, which bypassed Microsoft's RoguePlanet fix, allows a local attacker with limited permissions to reach SYSTEM privileges on affected Windows systems.

Source

ChromeOS 151 is rolling out to Chromebooks and ChromeOS Flex devices with 18 security fixes, real-time Safe Browsing for Gemini sessions, and a cutoff that leaves some older Flex PCs permanently on ChromeOS 150. The release also completes the transition announced when ChromeOS 150 ended legacy Chrome Apps in kiosk mode, so affected kiosk deployments must now use alternatives such as Progressive Web Apps.

Source

A newly disclosed Windows post-exploitation technique can activate the Chrome DevTools Protocol (CDP) inside an already running Google Chrome or Microsoft Edge process, giving attackers access to authenticated sessions without stealing or replaying cookies. The method requires prior code execution, but can expose browser data, WebAuthn activity, extensions, and live account access while bypassing protections aimed at off-device cookie theft.

Source

Microsoft has moved its rebuilt Print Management app into the Windows 11 Release Preview channel, putting the modern replacement much closer to general release. The decades-old console had previously appeared as an early experiment in a WinUI 3 Print Management rebuild; it is now included in Windows 11 builds 26100.9267 and 26200.9267, released as KB5120998 on August 14.

Source

GitHub’s latest Copilot update turns model choice into a core workflow rather than a one-time setting, adding Kimi K3, MAI-Code-1.1-Flash, and per-turn switching in Visual Studio Code. The flexibility comes with an enterprise concern: Copilot still does not provide a reliable record of which model generated or changed specific code.

Source

Anthropic’s planned Claude watermark does more than identify AI-generated text: it gives model makers another hidden influence over which words appear in every response. By subtly biasing token choices, providers can shape tone, emphasis, and meaning while claiming the changes are imperceptible. The technology risks turning private conversations, proofreading, and summaries into text influenced by secret provider-controlled rules that users cannot inspect or challenge.

Source

China’s 84% excitement and 72% trust in AI contrast sharply with the United States, where only 38% are excited and 32% trust the technology. The divide appears driven less by usage than by expectations: Chinese citizens are more likely to view AI as a practical productivity tool, while Americans focus on job losses, misinformation, privacy, and the power of technology companies. This difference in public attitude could become a decisive factor in the US-China AI race, because the eventual winner may be the country that can deploy AI most broadly across its economy—not merely build the most advanced models.

Source

Anthropic CEO Dario Amodei argues that open-weights models only partially decentralize AI because access to advanced chips and large-scale compute remains concentrated. That analysis understates how quickly hardware efficiency, quantization, and local deployment are expanding access: many useful AI workloads already run on laptops and smaller workstations without frontier-scale infrastructure. His position therefore appears less like a neutral assessment of decentralization than a defense of Anthropic’s business interests, since capable open models threaten the company’s pricing power, platform control, and role as an intermediary between users and AI capabilities.

Source

Anthropic explains that Claude’s text watermark works by changing the source of randomness used during word selection. When several words are similarly suitable, Claude uses a secret key and a short sequence of preceding words to choose among them. Each choice remains natural and meaning-preserving, but thousands of choices can form a statistically detectable pattern.

Source

Alibaba’s open-weight Qwen AI models have surpassed 3 billion downloads in the past six months—more than four times Google’s 418 million and Meta’s 227 million combined. The comparison is already a major win for Qwen, and it does not even include download figures for Kimi or DeepSeek.

Source

The Linux kernel is making the x86 Time Stamp Counter (TSC) mandatory, removing the final option to build a kernel without it. The change ends Linux’s decades-long compatibility path for very old processors and lets the kernel discard timing code that detected, calibrated, and worked around missing or unreliable TSC hardware.

Source

Anthropic has restored Claude after a roughly 36-minute outage disrupted authentication and degraded performance across Claude.ai, Claude Code, Claude Cowork, and platform.claude.com. The August 16 incident caused login failures, incomplete requests, and unresponsive interfaces, while Anthropic has not disclosed the root cause.

Source

Hazmat is a free, open-source containment tool for running AI coding agents without exposing an administrator’s normal user account, SSH keys, or cloud credentials. It creates a separate local account, limits access to a selected project, and wraps agents including Claude Code, Codex, OpenCode, and Cursor Agent.

Source

A new Microsoft Graph PowerShell workflow can permanently remove matching notification messages from Microsoft 365 mailboxes in batches of 20, reducing network round trips during tenant-wide cleanup. It extends the JSON batching approach for large Graph tasks from account updates to destructive mailbox operations.

Source

Linux 7.2 is now stable after a turbulent final stretch that followed an unusually large RC6 and an AI-heavy RC7. Linus Torvalds kept the August 16 release on schedule, despite a late surge of fixes and a rollback of problematic DRM GPU scheduling changes.

Source

Microsoft has moved the Windows Management Instrumentation Command-line (WMIC) removal into the KB5120998 Release Preview update, turning the planned WMIC removal into a concrete rollout. Windows 11 24H2 and 25H2 will no longer offer the legacy utility as a Feature on Demand, so administrators cannot reinstall it through Optional Features or DISM.

Source

Microsoft has issued a formal 60-day end-of-support warning for Windows 11 24H2 Home and Pro, Windows Server 2022, and Windows 10 Enterprise LTSB 2016. October 13, 2026, is the key deadline: Windows 11 and Windows 10 systems will stop receiving monthly security and quality updates, while Server 2022 will leave mainstream support.

Source