Anthropic’s expanded Cyber Verification Program includes Claude Mythos 5.1 alongside other models, with access organized into three tiers for security teams. The safeguards test described by Anthropic, however, used Claude Opus 5.5: it completed 34 of 50 cyber-operation challenges in Red Team Access without being blocked, while Defense Access blocked 46.

Source

Google’s EmbeddingGemma 2 can embed text, code, images, video, and audio in one shared space, enabling cross-media search without sending content to a cloud service. Its text-only configuration uses about 191 MB of active RAM, while the full multimodal model uses about 567 MB on Google’s tested phone. The Apache 2.0 model also supports smaller vectors to reduce local search-index storage.

Source

Underdog, a new invite-only AI assistant from Conway Research, runs on Macs and Windows PCs instead of sending prompts to the cloud. Founder Sigil Wen says it will be free and ad-free, with the company instead earning a small share of purchases made through the assistant.

Source

AI decision models offer a faster, more flexible way to apply content policies. Musubi’s open-weight PolicyLM-1.7B brings this approach to real-time moderation, applying plain-English rules to messages in under 50 milliseconds.

Source

AI agents such as Meta’s Muse are running into website blocks when they try to shop, book flights, or complete other tasks for users. Some blocks are deliberate; others come from anti-bot checks that mistake an agent for malicious automation. Meta, Walmart, Stripe, and other partners are working on an open standard to help businesses distinguish authorized agents from unwanted bots.

Source

Dell is adding a Unified Semantic Layer, Enterprise Knowledge Graph and Knowledge Agents to its AI Data Platform, aiming to give enterprise AI a consistent, governed view of company data. It is also adding NVIDIA cuDF GPU acceleration to its Data Processing Engine, with Dell reporting nearly four times faster processing on average than CPU-only systems. The context features are planned for the first half of 2027; the accelerated engine is scheduled for December 2026.

Source

Cline is an open-source AI coding assistant and agent platform that connects AI models to tools. A 4sysops review of Cline Desktop covers its session imports, scheduled agent jobs and model options. The company says token use on its platform has grown 20-fold in four months, with adoption rising after stronger open-weight models appeared in May. Cline is betting its open-source harness can help developers use those models for more than coding—and choose lower-cost inference when top-tier capability is unnecessary.

Source

Meta, Sierra, Stripe, Walmart and other partners plan to publish the first Personal Agent Protocol specification later this month. The initial version will focus on authenticating personal AI agents and controlling what they can access; payment support is planned for a later extension.

Source

Microsoft has defined four stages for the Azure Virtual Machines lifecycle, but hasn’t said how long VM series typically remain in each one. Customers will receive advance notice when older series approach end of life, and Microsoft says it will notify them directly about specific retirements.

Source

X.Org has patched 12 vulnerabilities in the X server and Xwayland; nine could allow arbitrary code execution. The fixes are in xorg-server 21.1.25 and xwayland 24.1.14, so check which version your systems run.

Source

OpenSSH 10.6, released October 6, makes the hybrid `ssh-mldsa44-ed25519` signature algorithm available without the vendor suffix used by its experimental predecessor. Administrators using keys from that earlier implementation must regenerate or remove them. The update also disables an SSH compression feature to address a newly described side-channel attack.

Source

Tuskira’s open-source AI Agent Gateway puts a control point between AI agents, their MCP tools, and model providers, keeping backend credentials out of agent configurations. It checks permissions on every tool call and can route model traffic while tracking token use and estimated cost. Before deploying it, bind keys to profiles and review its request-logging and network settings.

Source

A workaround using the Exchange Import-Export Graph API can retrieve folders and items from Microsoft 365 Group mailboxes—data the standard Outlook mailbox API cannot reach. It extends the known Graph blind spot for group mailbox folders, but it comes with extra setup and an important caveat: Microsoft’s published API scope does not include group mailboxes.

Source

Microsoft's September 2026 recap for Windows 365 and Azure Virtual Desktop (AVD) lists about two dozen changes. Several of them matter for daily administration: user-initiated provisioning for Windows 365 Reserve, business continuity and disaster recovery settings in Cloud PC configurations, local admin rights in Cloud PC configurations, display protection, TWAIN scanner redirection, and AVD Hybrid. Microsoft's recap is written like a marketing summary, so this article looks at the Microsoft Learn documentation to find out what you must configure and where the limits are. Some features are still in preview.

Source

Windows 11 now includes System Monitor (Sysmon), a tool that logs process creation, network connections, and file creation time changes to the Windows event log. Previously, it was a separate download from Sysinternals. The built-in version is an optional feature that is disabled by default, so you must enable it with PowerShell and install it with the sysmon command. This article explains the prerequisites, the commands, how to apply your own XML configuration, and how to verify that Sysmon writes events.

Source

The Dutch Tax and Customs Administration plans to move email and calendars from Microsoft 365’s cloud to on-premises servers, while seeking European open-source replacements for storage and collaboration by 2028. It is not a full Microsoft exit: Outlook and other Microsoft services will remain in use.

Source

Google’s Nano Banana 2.1 is rolling out with more precise mask-based edits and better consistency for people and objects across generated images. Developers can use the model through an API that accepts text, image, audio, and video inputs and generates images at up to 4K resolution.

Source

Atlassian is a software company that makes tools for managing projects, collaborating, and sharing knowledge at work. It is best known for Jira and Confluence, which help teams track what needs doing and document what they know. Atlassian and OpenAI are expanding their partnership to bring GPT-6 models, including GPT-6 Astra and GPT-5.6, to AI agents across Atlassian’s platform and Rovo. Rovo is Atlassian’s AI assistant, combining OpenAI models with the Teamwork Graph to draw on company context. Deeper Jira integrations that let teams assign work to agents are still being explored.

Source

The top 100 gen AI consumer apps are ranked by spending, mobile users, and web visits: ChatGPT leads both mobile and web traffic, while OpenAI tops consumer revenue, ahead of Anthropic and Canva.

Source