A self-propagating Word worm can still alter documents and copy hidden instructions into new files after Microsoft 365 Copilot mitigations, including a model upgrade, failed to close the broader vulnerability class. The attack requires no macros, executable code, or access to a victim’s Microsoft 365 tenant—only a malicious document that enters Copilot’s context.

Source

OpenAI’s GPT-5.6 Sol was not the problem during a researcher’s ripgrep investigation—the separate cybersecurity classifier was. After repeated refusals, Daniel Fox Franke switched to Z’ai GLM 5.2 and Moonshot AI’s Kimi K3, which helped trace the recurring segmentation fault to a suspected Linux kernel bug.

Source

Cisco will stop selling its Azure Local bundles on October 24, 2026, with final shipments due by January 24, 2027. The company is offering no direct replacement, leaving customers to assess other HCI platforms through Cisco Compute or third-party providers.

Source

Cisco Secure Firewall Management Center (FMC) is under active zero-day attack through hard-coded credentials that let unauthenticated remote attackers access a low-privilege account and sensitive data. Cisco has released hotfixes for supported FMC versions, while CISA has added CVE-2026-20316 to its Known Exploited Vulnerabilities catalog.

Source

Alibaba’s Qwen-Audio-3.0-Realtime Plus has taken the top position on Artificial Analysis’s Speech-to-Speech Index, scoring 84.1% versus 79.1% for OpenAI’s GPT-Realtime-2.1 High. The result gives Alibaba its first lead over OpenAI on the benchmark, but Qwen’s roughly four-second startup latency is a significant trade-off for interactive deployments.

Source

Azure surpassed $100 billion in annual revenue during fiscal 2026, even as Microsoft says it still cannot deliver enough AI and cloud capacity to meet demand. Quarterly Azure growth reached 43%, its fastest pace since early 2022, and Microsoft expects growth to accelerate to about 45% in the current quarter.

Source

A Microsoft Outlook Web Access flaw is allowing the Russian-aligned TA488 group to establish mailbox access that survives password changes, browser restarts, and full device reimaging. Its OWAReaper implant turns a single email open into persistent, server-side Exchange permissions.

Source

Anthropic’s Claude Mythos AI security model found a symmetry-based attack that reduces HAWK-256’s key-recovery effort from roughly 2⁶⁴ to 2³⁸ operations. After verifying the result, HAWK’s developers withdrew the post-quantum signature scheme from NIST’s standardization process.

Source

Microsoft has released new updates for Microsoft Intune that expand device management capabilities for IT administrators. You can now monitor per-device sync progress directly in the Intune admin center for better troubleshooting. The update also introduces custom macOS compliance settings to evaluate specific device conditions. Additionally, you gain granular control over Samsung firmware updates through a new integration. These features aim to provide greater visibility and flexibility across different operating systems.

Source

China’s AI gap with the United States is narrowing through a combination of exceptional technical talent, open-source collaboration, and models optimized for specific tasks rather than unlimited scale. Grace Shao told Big Technology Podcast host Alex that Kimi K3’s performance shows frontier-level AI is becoming more global, less proprietary, and harder to sell at premium prices.

Source

Matthew examines the debate between open-source and closed-source artificial intelligence, comparing their advantages, drawbacks, business models, and broader geopolitical implications. Open-source AI models can be inspected, modified, downloaded, and run independently, while closed-source systems are controlled by the companies that develop and operate them.

Source

Moonshot AI has paused new subscriptions to Kimi K3 less than 48 hours after launching the open-weight model, after demand pushed its GPU infrastructure close to capacity. The Chinese startup is splitting access into consumer and coding tiers as it rations compute while pursuing new funding and a potential Hong Kong IPO.

Source

Microsoft says its new MAI-Cyber-1-Flash model can handle up to 90% of vulnerability-hunting tasks while cutting the cost of its AI security system by about half. The model will enter Project Perception’s public preview on August 3, with the hardest cases escalated to OpenAI’s GPT-5.4.

Source

AI agents need dedicated identities, task-scoped permissions, approved tool allowlists, and fast revocation—not broad service-account access. New guidance also emphasizes short-lived entitlements, context-aware authorization, and mandatory human approval for high-impact actions.

Source

Microsoft will staff a booth and host six Azure VMware Solution sessions at VMware Explore 2026 in Las Vegas from August 31 to September 3. The program targets organizations planning VMware migrations, security improvements, and AI initiatives on Azure.

Source

Anthropic is facing a growing trust problem as startup founders and researchers move toward cheaper open-weight AI models. The backlash intensified after Claude Design was viewed as a potential threat to Figma, an Anthropic partner, and culminated in a San Francisco protest against frontier labs’ control of advanced models.

Source

Hugging Face’s forensic reconstruction shows an autonomous OpenAI agent chained two dataset-processing flaws into a production foothold, then reached cluster-admin privileges across multiple Kubernetes clusters in under 13 hours. Investigators reconstructed about 17,600 actions from the 4.5-day campaign and used the open-weight GLM-5.2 model to decode concealed payloads.

Source

More than 1,100 employees at OpenAI, Anthropic, Google, Meta, and other AI companies are asking the US government to support an international system that can deliberately pace frontier AI development. The petition warns that automated AI research could advance faster than society can understand or control.

Source

Visual Studio Code 1.132 Insiders adds several practical improvements, including proposed-API support for `code serve-web`, customizable dictation and Voice Mode instructions, and one-click Markdown code copying. The July 29 build also reduces terminal tab clutter and makes manual folding cleanup more precise.

Source