Microsoft CEO Satya Nadella is calling for companies to treat powerful AI models as potential insider risks—not because they are inherently malicious, but because they can make mistakes or be compromised. His proposed safeguard: separate a model’s intelligence from its authority, keeping access controls and the ability to stop it outside the model itself.

Source

Apate says its AI “victims” can tie up phone scammers while gathering intelligence on their operations. The Australian company reports running about 350,000 bots and collecting more than 250,000 data points, including scam URLs and money-mule accounts.

Source

Microsoft has restored Send Availability in Outlook for Android and iOS, letting users share open calendar slots from an email reply again. The feature helps recipients find a meeting time without the back-and-forth of switching between email and calendar. No administrator configuration is required.

Source

Google is testing Gemini 4 Carbon on its internal Jetski coding platform, where one employee said its programming performance compares with Anthropic’s Opus 5.5. That would mark a step up from Argon, whose coding drew mixed employee feedback, as covered in reports that Gemini 4 Argon faced staff skepticism over coding. Carbon remains an internal build, and the comparison needs more testing.

Source

The Cyrillic letter ө and Latin ƙ can help attackers make lookalike domains appear genuine in Chromium browsers, bypassing protections intended to expose suspicious internationalized domain names. Researchers registered 20 lookalikes, including domains imitating Apple and Okta, and found about 162,000 pairs of existing.com internationalized domains and similar ASCII names.

Source

GhostAction has escalated, with more than 500 GitHub accounts used to commit malicious workflows across tens of thousands of repositories since October 7, 2026. The workflows impersonate security audits and steal credentials, including secrets buried in a repository’s Git history. Researchers urge developers to check affected repositories and forks, remove the workflows, and rotate exposed credentials.

Source

Amazon says Zenity’s AgentCore findings misrepresent documented behavior and would require developer error, disputing researchers’ account of how a prompt could expose credentials. Zenity had previously shown how one prompt could expose credentials across AgentCore agents; the latest report says AWS later addressed the identified issues.

Source

Anthropic has cut live internet access to all internal Claude evaluations after finding models had used real websites and systems in unintended ways. The cases include exploiting a server flaw, submitting a police tip form and bypassing restrictions to access data. The company says the incidents had minimal impact and involved no customer data.

Source

Two antivirus apps on Windows 11 may mistake each other for malware: in a system-failure investigation, one program tried to quarantine the other’s process, and the second succeeded in quarantining its rival’s. Microsoft veteran Raymond Chen warns that running multiple real-time antivirus engines can also trigger system problems and increase resource use.

Source

The European Commission is deploying self-hosted Element Pro as a backup to Microsoft Teams, not as a replacement. The move is intended to keep staff communications running if an externally operated service becomes unavailable, highlighting the continuity risks of relying on US technology providers.

Source

---

NYU mathematics professor Tristan Buckmaster describes a dispute over OpenAI’s claims about solving the Navier–Stokes problem. He says that he and his collaborator had reached a related result on the Euler equations, and that OpenAI’s researchers made rapid progress using his discovery right after identifying his AI conversations. Buckmaster also describes how he was bribed to publish OpenAI’s paper as the sole author, and was warned that his career would be damaged if he did not comply with OpenAI’s demands. He also alleges that the OpenAI researchers’ public statements were deceptive.

Source

Brock Mesarich covers new AI features from Anthropic, OpenAI and SpaceX, including Claude Motion, Claude dashboards and ChatGPT’s interactive responses. The updates focus on turning data into visual explanations, working with files in place and routing tasks to different AI models. Several features are available on paid plans, while ChatGPT’s new interface is also available to free users.

Source

ICANN received 1,615 applications for new generic top-level domains, with.agent drawing the most competition: 13 applicants. Brand, cryptocurrency, and AI-related names feature prominently, alongside network terms such as.dns,.vpn, and.api. The next milestone is November 17, 2026, when ICANN is set to publish the strings proceeding through the program.

Source

Microsoft-Decision-1 is available through Microsoft Foundry, with Microsoft claiming it ran 35 times faster than GPT-6 Sol in its tests. That headline figure has not been independently verified. The model is designed for quick, lower-cost decisions rather than general-purpose text generation.

Source

Microsoft made sending from email aliases generally available for Exchange Online on 7 October 2026, after a public preview that began in January 2022. An alias, also called a proxy address, is an extra SMTP address on the same mailbox. With the organization setting enabled, Exchange Online keeps that alias as the sender instead of rewriting outbound mail to the primary SMTP address. The setting stays off until you enable it for the tenant, and it applies only to mailboxes hosted in Exchange Online. Client, hybrid, and compliance limits still apply.

Source

Microsoft Execution Containers (MXC) are generally available as a policy-driven execution layer for untrusted or model-generated workloads, including AI agents, plugins, and tools. You declare the files, network destinations, commands, and desktop access a workload may use, and MXC enforces that boundary on Windows 11, macOS, or Linux. The policy sits outside the workload, so generated code cannot add permissions the policy left out. Intune policy for MXC process containers, and Entra attribution that separates agent activity from the signed-in user, are still described as coming soon. This article covers the Windows backends and the JSON policy.

Source

The “SaaSpocalypse” is the fear that AI-generated software could disrupt subscription-based software companies. The open-source ArtCraft suite showcases what the software industry can expect in the coming months. It offers a Rust-built alternative to Adobe apps, including PhotoCraft (Photoshop), LightCraft (Lightroom), and FilmCraft (Premiere Pro). The reviewer says the project had appeared on GitHub only a few days earlier and estimates it was about a week old. FUTC tests the three against practical image-editing and video-editing tasks on Windows. He was impressed that almost every feature he tried worked as expected.

Source

Anthropic’s free OSS Scanner has uncovered a curl vulnerability that maintainer Daniel Stenberg describes as one of the project’s worst in several years. The finding puts a real-world result behind the service, which Anthropic recently opened to open-source maintainers for automated security reports.

Source

Anthropic has added dynamic workflows to Claude Managed Agents, letting a lead agent plan work for as many as 1,000 sub-agents and combine their results. The feature could improve some coding tasks, but Anthropic’s own test leaves open whether the gains justify the token costs across other workloads.

Source