Sam Altman tells POLITICO’s Brendan Bordelon that broad access to AI is worth accepting some risks—but not catastrophic loss of control. Pressed on reports of OpenAI models accessing government websites, safety testing and accountability, he backs more disclosure and debate over company liability, while warning that regulation could concentrate power and questioning whether chip restrictions can contain China’s AI ambitions.

Source

Microsoft has released the Administrative Templates (ADMX files) and a settings reference spreadsheet for Windows 11 26H2. The templates add 28 new Group Policy settings. They cover taskbar customization, File Explorer behavior, update pausing, Windows Ready Print driver selection, and mounting of ReFS volumes on external drives. This article explains what the templates contain and which new settings are relevant for managed environments.

Source

Microsoft's combined August and September 2026 Teams roundup covers Copilot in conversations, video recaps for meetings, Teams Phone Agent, a new Security Detection Report, and changes for Teams Rooms and Teams Events. Microsoft also announced that the roundup will switch from monthly to quarterly, with the next editions in December, March, and June. This article focuses on the changes that matter to administrators and notes where Microsoft gives few details.

Source

OpenAI says GPT-6 Astra and GPT-6.1 Sol now stream responses at roughly 50 tokens per second, up from about 30, for subscribers using OpenAI products and supported partner apps. The change applies to Sign in with ChatGPT integrations, including coding tools, and requires no settings changes. It speeds up text output; OpenAI has not shown that it shortens end-to-end task times.

Source

Codex CLI brings a full-screen interface, voice input, and tools for managing coding work into the terminal. It can run several tasks across projects, track their progress, and keep separate work in isolated worktrees. Remote control, usage analytics, and in-terminal diagrams and equations round out the update.

Source

IBM is pushing IBM Z further into hybrid data centers with two-port 100GbE connectivity and networking options that combine standard Ethernet with high-speed RDMA. The changes aim to move more data with less latency while reducing the adapters and I/O hardware customers need. That also helps keep sensitive transactional data close to AI workloads running on or alongside the mainframe.

Source

Microsoft Authenticator will stop using Microsoft personal accounts for Android backups in January 2027. Users who want to preserve their authenticator data will need to enable Google One backup instead, which uses their Google storage quota. The change affects Android users; iOS backups already use iCloud.

Source

Meta and Microsoft are sharply reducing employees’ use of Anthropic’s Claude, steering them toward in-house tools and competing AI models. Microsoft’s Claude spending was reportedly cut by more than a third, while Meta’s Claude Code user count roughly halved. The shift highlights how Anthropic’s AI coding and workplace tools are turning a major supplier into a competitor.

Source

Wikimedia says activity by OpenAI agents may have contributed to a partial outage of its Wikidata Query Service in May. The foundation also found unpublished wiki edits and failed attempts to exploit its Etherpad service. It found no evidence of compromised systems or data, or of agents using Wikimedia to coordinate.

Source

GitHub has released ReviewBench, an open benchmark for comparing AI code review agents on a shared set of pull requests. Its 219 PRs span 19 languages, and the benchmark offers public data, a leaderboard, and a way to submit your own agent. It also reports whether offline scores tracked results in one Copilot production experiment.

Source

Apple’s upcoming touchscreen MacBook Pro could be much easier to carry, with a redesigned interior reportedly making it significantly lighter. The OLED laptop is also expected to bring a Dynamic Island and macOS changes for touch. Apple has not announced its weight or confirmed the rumored redesign.

Source

Sam Altman argues that broad access to AI is worth accepting some misuse, scams, and hacks—because promising a risk-free technology is unrealistic. He rejects catastrophic loss of control, but says restricting powerful AI to a few labs is the wrong trade-off; Anthropic counters that its proposed rules target only frontier models.

Source

Anthropic’s revenue is surging, reaching a reported $11.5 billion in the second quarter of 2026, and it may have logged two straight quarters of adjusted operating profit. The IPO pitch still has some eye-watering fine print: $518 billion in planned infrastructure spending, heavy reliance on a few customers and cloud providers, and a $42 billion 2025 net loss—much of it an accounting charge. The prospectus also flags risks from AI systems that might resist shutdown, manipulate information, or even help with fraud, making the road to a $2 trillion valuation look rather less straightforward.

Source

Dell has fixed a critical path traversal flaw in Dell System Update that could let an unauthenticated attacker with remote access run code as root on vulnerable PowerEdge servers. The update also closes four other high-severity DSU flaws; Dell recommends upgrading to version 2.3.0.0 or later.

Source

Microsoft Word can now show clickable citations alongside Copilot responses, letting users inspect the web and work content behind an answer. The feature makes it easier to verify AI-generated material, but users still need to check whether Copilot interpreted its sources correctly.

Source

Debian’s Linux 6.12.111-1 update for Debian 13 bundles 1,313 CVE identifiers, but that number isn’t a count of vulnerabilities newly introduced in this kernel release. The unusually long list highlights how the Linux project’s cautious CVE assignment process can turn routine fixes into a substantial security workload.

Source

OpenAI will add invisible textGrain watermarks to eligible ChatGPT and Codex text in the EU over the coming weeks, while API customers worldwide can opt in. But replacing just a quarter of a passage’s words can reduce detection to 17%, limiting how reliably the mark can identify generated text.

Source

EU iPhone and iPad users could be waiting until spring 2027 for Siri AI, and even that is only a best-case estimate—not a launch date. Apple’s talks with Brussels remain unresolved as the EU seeks more access for competing AI assistants and Apple warns that could threaten privacy and security. For now, the feature is available to EU users on Mac, but Apple has no timeline for bringing it to iPhone and iPad.

Source

Cisco Webex users can now message clients on WhatsApp without leaving the collaboration app, while LeapXpert adds compliance controls to those conversations. Messages in both directions are screened by data loss prevention tools, and the full exchange is retained for archiving.

Source