Matthew Berman calls OpenAI’s GPT-6 Astra his most capable model yet, highlighting major gains in coding, spatial reasoning, browser control, and autonomous game creation. His demonstrations show Astra building functional 3D environments from short prompts, while benchmark results and safety testing suggest a substantial advance over GPT-5.6 and Claude Fable 5.1.

Source

Nvidia CEO Jensen Huang and Hugging Face CEO Clem Delangue discussed Nvidia’s $12.93 billion acquisition of the open AI platform, saying the deal will accelerate access to models that organizations can run, adapt, and control themselves. Hugging Face will remain an independent, neutral platform within Nvidia, with its founders and team joining the chipmaker.

Source

Google is pushing a more flexible pricing model for AI agents as companies struggle to control rapidly rising usage costs. The plan combines pay-as-you-go billing, token discounts of up to 20%, hard monthly spending caps, and a zero-cost base subscription—directly challenging Anthropic’s recurring seat fees and Microsoft’s separate usage licenses.

Source

Microsoft's August 2026 security roundup covers managed detection for third-party logs, multi-tenant identity governance, Intune device onboarding and remote support, higher Purview auto-labeling throughput, and new guidance for containing AI agents. If you manage Windows endpoints and Microsoft 365, these changes affect how you onboard devices, label data before using Copilot, and prepare for autonomous software agents.

Source

AI-assisted development is moving beyond one-off prompts toward repeatable workflows built from loops, specialized agent teams, and the tooling that controls them. For system administrators, the emerging vocabulary matters because it describes how AI systems will access code, tools, permissions, infrastructure, and validation processes.

Source

Amazon Quick now connects directly to Microsoft Outlook, allowing users to summarize threads, draft contextual replies, manage messages, and schedule meetings without leaving the mailbox. The extension can also trigger broader workflows, but its ability to take actions makes permissions, data handling, and response verification important considerations for IT teams.

Source

AWS has published an enterprise deployment pattern that routes OpenAI ChatGPT Codex through LiteLLM on Amazon ECS and Amazon Bedrock, giving teams centralized model access, budgets, rate limits, and usage visibility. The important detail is that Codex still runs its task loop and approved tools on the developer’s workstation; only model requests pass through the AWS-hosted gateway.

Source

GPT-6 Astra is now rolling out to selected organizations, and OpenAI President Greg Brockman says the model may mark the beginning of the “AGI era.” The release follows Astra’s designation as OpenAI’s first model to reach the “critical” cybersecurity capability threshold, meaning it can autonomously discover unknown vulnerabilities and construct exploit chains.

Source

A corporate password found in an infostealer log should be treated as a possible live identity compromise, not just an old credential leak. Security teams need to determine within minutes whether the data includes active browser sessions, identity-provider access, VPN or RDP credentials, and then revoke sessions and reset passwords before attackers can use them.

Source

Spring Ring attackers used fake Microsoft Teams support accounts to target more than 150 employees at at least 10 organizations, turning live calls into a way to deploy malware or pursue NTLM relay attacks. The campaign ran from January through April 2026 and shows why unsolicited Teams conversations deserve the same scrutiny as email phishing.

Source

Nvidia’s free, open-source Personal AI Router (PAIR) can combine idle PCs, Macs, and DGX Spark systems on a home or small-office network to process local AI workloads. The beta routes inference requests to available machines, potentially delivering up to 1.9x faster local inference while keeping prompts and data off cloud services.

Source

Google is rolling out Gmail Live, Docs Live, and Keep Live, allowing users to search email, create documents, and manage notes through spoken conversations with Gemini. The English-language features are now generally available on mobile for eligible Google AI subscribers, while Google Workspace business customers are still waiting for access.

Source

Microsoft is adding a human-approval gate to Copilot Studio, allowing organizations to require explicit permission before an AI agent or one of its tools takes action. The safeguard is aimed at preventing autonomous agents from sending emails, closing tickets, processing payments, or changing sensitive data without human oversight.

Source

Microsoft is challenging OpenAI, Google, and ElevenLabs with MAI-Transcribe-2, a speech-to-text model that supports 60 languages and costs just $0.10 per hour of audio. Microsoft says it combines near-frontier accuracy with processing speeds up to 10 times faster than competing models, although the introductory price is only guaranteed through the end of 2026.

Source

AWS has expanded Amazon Elastic VMware Service (EVS) to 22 regions and doubled the supported environment size to 32 hosts, giving organizations a broader path for moving VMware workloads out of aging data centers. The service now supports VMware Cloud Foundation (VCF) 9.0 and 9.1 on Amazon EC2 bare-metal infrastructure, including a lift-and-shift VCF-to-VCF migration model that preserves existing VMware operations.

Source

CrowdStrike has introduced its Agentic Identity Provider, a system designed to identify and manage autonomous AI agents before they receive access to enterprise resources. Announced at Fal.Con 2026, the service uses Falcon Guardian to discover agents as they appear and replaces improvised service accounts and API keys with identities that can be evaluated continuously.

Source

BraZetsu is turning infected Windows computers into inventory for Exilware’s underground access marketplace, with AI-assisted profiling used to estimate what each foothold is worth. The Python-based framework primarily targets Brazil, Iberia, and Latin America, but operators were also seen advertising access to U.S. systems in April 2026.

Source

Cisco has patched CVE-2026-20212, a critical Nexus 9000 vulnerability rated 9.8 that lets an unauthenticated remote attacker execute code as root through TCP ports 43210 or 43211. The flaw affects 10 Silicon One-based switch models and can also crash the S1HAL process, forcing a device reload and potentially disrupting data-center traffic.

Source

Microsoft is turning its Driver Quality Initiative (DQI) into more visible enforcement, with stricter driver-signing checks, cloud-based rollback, and a pilot designed to deliver display drivers only to compatible systems. Windows 11 quality updates expected in October may also enable Memory Integrity on more eligible PCs, improving protection against malicious or untrusted kernel drivers but potentially reducing gaming performance on older hardware.

Source