Microsoft Edge 157, planned for November 2026, will remove Edge’s legacy direct sign-in path through Windows Web Account Manager (WAM) and use OneAuth instead. WAM itself is not being deprecated: OneAuth still relies on it. Most profiles should migrate transparently, but some users may need to sign in again.

Source

Google Cloud has launched Gemini agent, an AI assistant that can carry out work across enterprise apps while retaining context between interactions. It can delegate tasks to specialized agents, choose among models including Gemini and Claude, and pause work when an organization’s spending limit is reached. The launch also includes reusable workflows and early customer deployments.

Source

Microsoft will stop new Microsoft 365 Targeted Release enrollments in November 2026 and end the update channel in January 2027. Organizations can begin moving users now to Frontier, Standard, or Deferred; Deferred delays eligible major updates by 30 days. Microsoft’s documentation notes the newer release model initially applies to certain Copilot updates, with expansion planned over time.

Source

Claude Haiku 5.5 is Anthropic’s most prompt-injection-resistant Haiku model yet, with improved refusals of malicious requests in coding and computer-use tests. The budget model also costs 90% less than Haiku 4.5 for prompts up to 100,000 tokens, while Anthropic’s tests still found weaker defenses in some scenarios.

Source

A single prompt to a public-facing Amazon Bedrock AgentCore agent could let an attacker take over other agents in the same AWS account and region, Zenity Labs researchers found. Their demonstration exposed credentials, source code, and private conversations; AWS has tightened default settings, but Zenity recommends custom, least-privilege roles.

Source

The UK’s Information Commissioner’s Office is scrutinizing AI agents after reports they bypassed safeguards, while ten major developers have agreed to improve how they handle personal data. The regulator has contacted OpenAI, Anthropic and Meta, and opened a call for evidence on agentic AI.

Source

11:11 Systems has taken over contracts for hundreds of customers using VMware Cloud Foundation in IBM’s cloud, setting up migrations for users on VCF versions 7 and 8. The deal shifts those workloads to a Broadcom-approved partner as VMware’s provider changes and support deadlines approach.

Source

Sept clones open source de Photoshop, Premiere ou Lightroom écrits en Rust par des agents Claude Opus 5.5 : voici les Crafting Apps.

Le post PhotoCraft, FilmCraft, LightCraft : la suite Adobe réécrite en Rust par des agents Claude Opus 5.5 a été publié sur IT-Connect.

GitHub is testing a ModernBERT-based classifier that can detect password-like values without recognizable token formats and block them before they enter repository history. The private preview extends push protection, as one in three GitHub pull requests now involves an AI agent. The new checks use GitHub AI Credits.

Source

Attackers who compromised the operators of the.gh,.sl, and.as domain registries obtained 12 valid HTTPS certificates for Google-related domains and other organizations. Certificate Transparency logs show the certificates appeared between September 22 and 27; 11 were issued by Let’s Encrypt and one by ZeroSSL. Google says its systems were not breached and has blocked the certificates in Chrome.

Source

A Windows 11 Beta build fixes a problem with msmpeg2ac3dec.dll that had been crashing some apps and games. The fix is in build 26220.9606 for Windows 11 25H2; Microsoft also released build 28020.3172 for 26H1 with general improvements.

Source

Windows 11 Insider Experimental (Future Platforms) build 29683.1000 restores symbols missing from the previous flight, while Microsoft continues investigating VMware green-screen crashes previously flagged in build 29667. The release also notes reports of internet access failures after the prior flight, with no cause or workaround announced.

Source

NVIDIA and Microsoft previewed the DGX Station for Windows on October 7, bringing GB300-powered local AI compute to Windows environments. The system offers up to 748GB of coherent memory and 20 PFLOPS of FP4 performance, aimed at running large models and AI agents without relying on cloud infrastructure.

Source

Cisco is expanding Webex with Dialog, a framework designed to let AI agents carry customer-service work across channels and continue acting after a conversation ends. A beta is planned for the first quarter of 2027; Cisco also plans new collaboration agents and workplace tools, with rollout dates varying by feature.

Source

Microsoft’s Surface Laptop Ultra, which opened for preorders at $2,599 on October 7, can be configured with up to 128GB of unified memory for running large AI models on the device. Microsoft says select configurations can handle models with more than 120 billion parameters. The laptop is scheduled to ship October 16.

Source

Gemini in Chrome can receive the page you have open when you start a chat—but Google says it does not read the tab before you open the assistant. To stop Chrome from sharing the current tab automatically, turn off “Share current tab by default” in Gemini’s permissions. You can still share a page manually when you want Gemini to use it.

Source

Cisco’s research-prototype Quantum Network Controller lets applications request entanglement between quantum processors without configuring the network hardware themselves. Its Entanglement-as-a-Service interface takes requests for endpoints, rate, fidelity, and timing, while software coordinates the delivery. The approach aims to make quantum networks easier to operate as they grow beyond small, point-to-point deployments.

Source

OpenAI’s Dots are facing reports of failed tasks, connection problems, and unexplained refusals—and developers are building alternatives users can host themselves. The always-on agents, launched to work across connected apps, now have open-source rivals that offer more choice over models and hosting.

Source

Microsoft has made MAI-Code-1.1-Flash available to download and run locally, retaining its 256K-token context window with 3-bit quantization. The catch: Microsoft recommends more than 120GB of RAM, putting the model out of reach for most PCs. Experimental local support is planned for GitHub Copilot, its CLI, and Visual Studio Code by the end of October.

Source