HOL Guard is an open-source security layer that pauses risky actions from AI coding agents before they reach the host system. The tool runs locally, works offline, checks requests in under 50 milliseconds, and has passed 552,000 downloads—but its developers cannot yet say how many users keep it enabled.

Source

Google’s Gemini Enterprise AI stack combines the Gemini Enterprise app, Gemini agents, Gemini Enterprise Agent Platform, Gemini Code Assist, CodeMender, the Gemini API, Gemini models, Spark, and Apigee, with MCP and A2A providing standardized ways for agents to connect to tools and each other. These are separate products and protocols rather than one installable platform: some provide models or AI applications, while others support development, automation, API management, or agent communication.

Source

Leaked Microsoft documents and video reveal Project Aion, a working Copilot-only operating system that replaced the Windows desktop with an AI-driven web shell. Designed for PCs and potentially Android-based devices, Aion abandoned the Start menu and Win32 apps in favor of task-focused conversations, while Microsoft’s current priorities suggest the project is unlikely to ship.

Source

Cloudera’s global survey finds that 95% of enterprises postponed or canceled AI initiatives during the past year, while 72% say their data architecture needs a major redesign. The findings show how governance, compliance, rising costs, and increasingly distributed data are pushing organizations toward a broad infrastructure reset.

Source

OpenAI’s GPT-5.6 Terra and Luna models are now generally available through Amazon Bedrock in both AWS GovCloud (US-East) and AWS GovCloud (US-West), bringing advanced AI workloads to AWS’s isolated government cloud regions. The release combines a choice between balanced and fast inference with a one-million-token context window, 90% discounted prompt caching, and access through Bedrock’s standard tooling.

Source

ChatGPT Search sharply reduced its Reddit citations in August 2026, with Reddit’s share falling from 3.83% to 0.52% in Promptwatch’s tracking. The timing coincides with ChatGPT greatly increasing domain-specific `site:` searches and nearly doubling the number of web queries it performs for each answer.

Source

Nous Research has bundled Bot Mode into Hermes Desktop and enabled it by default in Hermes Agent v0.20.3, moving the feature from a one-day public beta into the main product. It turns each Hermes profile into a named bot with separate memory, credentials, skills, chat history, and model settings, then lets bots hand work to one another.

Source

Canonical is funding a three-year University of Bristol research project to determine whether AI can convert large, mature C codebases into safe, maintainable Rust—not to immediately rewrite Ubuntu. The project will test a hybrid approach combining machine learning, program analysis, and formal methods on AppArmor and snap-confine.

Source

IBM is developing a dual-architecture processor with Arm that will let the same core run z/OS and Arm-native Linux workloads concurrently on future IBM Z and LinuxONE systems. Announced at Hot Chips 2026, the design could give regulated organizations direct access to Arm’s AI and cloud-native software ecosystem without moving sensitive workloads off their mainframes.

Source

Chrome is testing “Critical Actions in History,” a new Chrome Canary feature that records what Gemini does during browser tasks. The audit view follows Google’s experiments with letting Gemini use saved passwords to complete web tasks, adding visibility into actions such as password fills, file downloads, and visited pages.

Source

OpenAI is temporarily cutting GPT-5.6 Sol API pricing by up to 33.3%, lowering generated-output costs from $30 to $20 per million tokens through November 21, 2026. The promotion follows July’s GPT-5.6 repricing, which reduced Terra and Luna prices but left Sol at its original rates.

Source

Microsoft is rolling out a Teams meeting-policy control that detects external recording and AI notetaking bots and places them in the lobby for approval. Targeted-release tenants are receiving the feature now, with worldwide availability expected by the end of September 2026, giving administrators a way to prevent unapproved applications from entering confidential meetings.

Source

Anthropic identified the cause of a new Claude outage within 21 minutes, but developers still faced elevated errors while the company applied a fix. The August 24 incident affected Claude Mythos 5, Fable 5, Opus 5 and Opus 4.8, along with Claude.ai, the API, Claude Code and Claude Cowork.

Source

The new Model Context Protocol (MCP) roadmap shifts attention from the recently released stateless architecture to the next problems of running AI agents at scale: server-initiated events, agent identity, progressive tool discovery, and clearer tool results. It builds on MCP’s stateless 2026-07-28 release, which removed protocol-level sessions and made remote MCP servers easier to operate behind standard HTTP infrastructure.

Source

Two years after Salesforce launched Agentforce, its partners report growing customer curiosity but no meaningful revenue from the AI platform. Salesforce partners are consulting firms, technology providers, resellers, and implementation specialists that help customers adopt Salesforce products. A TD Cowen survey found that early trials and buying activity have not yet turned Agentforce into a source of partner bookings, challenging Salesforce’s claims that demand is accelerating.

Source

Nvidia’s Agentic Variation Operators (AVO) is a research harness that wraps an AI model with tools, memory, feedback loops, and a supervisory agent. By placing Claude Opus 5 inside this system, Nvidia lifted its score from 30% to 100% on ARC-AGI-3, solving all 183 puzzles across the benchmark’s 25 interactive games. The result suggests that harness design can matter as much as, or more than, changing the underlying model for difficult, long-running agent tasks.

Source

Anthropic is expanding Claude Mythos 5 beyond its limited defender program: Claude Enterprise customers can now use it through Claude Security to scan codebases, identify vulnerabilities, and propose patches. The company is also integrating the model into partner security products and launching a $35 million credit fund for open-source software security, while keeping direct model access restricted.

Source

DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model that can interpret images and screenshots and act on what it sees. The company says it keeps the text performance of V4 Flash while approaching Anthropic’s Opus 4.8 on multimodal agent tests.

Source

Anthropic’s new invisible watermarks for Claude-generated text are already being circumvented in multiple ways. Within hours of the announcement, Guillaume Meyer published an open-source remover, while other developers demonstrated rewriting, character substitution, sentence reshuffling, and translation-based techniques that can erase or weaken the machine-readable signal.

Source