Today in AI — October 7, 2026

OpenAI and Anthropic traded headline launches earlier today, both covered as they landed; the rest of October 7 was about where AI actually runs — on your laptop, inside the rack, across the domain registry — and what else shows up on those machines, up to and including malware taking orders from a poem.

Models & Research

• A 12-model panel found that Chinese models are the biggest conformists. A small tool that asks twelve models for their independent top-three picks and scores each against the consensus of the other eleven put DeepSeek, Kimi K2, Tencent's Hy3 and Zhipu's GLM in the top four; the five Chinese models averaged 0.47 versus 0.38 for the six American ones, with Claude Haiku the most contrarian at 0.32. The author flags the obvious caveat — a five-model Chinese bloc helps define the consensus it is measured against — but the distillation explanation lands on the labs most often accused of it, and the deeper worry is real: consensus that gets gamed once can be copied until it looks like a fact about the world.
• Microsoft's bug-hunting agents found 140 Windows CVEs in five months — and proved the bottleneck is no longer discovery. From May through September, the FORGE Lab's agentic research fed 140 Windows vulnerabilities into Patch Tuesday (52 in September's release alone) plus 155 validated reports across 23 open-source projects, 93 of them already acknowledged or accepted by maintainers including curl, FFmpeg, Linux, llama.cpp, Node.js, vLLM and SQLite. Producing a proof of concept for a confirmed Linux kernel crash averaged $3.61 in model cost and about 21 minutes on GPT-5.5; Microsoft's lesson for security teams is to count validated fixes, not findings, because once agents can find hard bugs, review and remediation become the scarce resource.

Industry

• Microsoft and Nvidia put local AI on sale: RTX Spark laptops from $2,599, a developer box for $5,999. The Surface Laptop Ultra ships October 16, while the Surface RTX Spark Dev Box — 128GB of unified memory, aimed at running local models beyond 120B parameters — follows in November; first-wave RTX Spark machines from Dell, Lenovo, MSI, HP and Asus stretch to about $7,000 for maxed configs. On stage with Satya Nadella and Jensen Huang, Microsoft also gave Copilot "Hybrid Intelligence": access to local files and cross-app actions, demonstrated with an agent that found tax documents, zipped them and drafted the email — and confirmed Meta's Muse is coming to Windows.
• Muse reached the iPad exactly one month after phones. Meta's agent now has a dedicated iPad app — a milestone Instagram itself took about 15 years to reach — and has racked up roughly 6.6 million installs since its September 8 mobile debut, per Sensor Tower estimates. The release also added a batch of connectors spanning Asana, Canva, Figma, GitHub, QuickBooks, Notion and Zoom, the plumbing that turns a chat window into something that can actually do your errands.
• Former White House AI adviser Sriram Krishnan is raising about $500 million. According to The Information, the fund will back growth- and later-stage US AI companies with national security as the priority — and Krishnan was on Microsoft's stage today, which tells you how close the policy-to-venture revolving door currently spins.
• Arista is pushing Ethernet inside the rack, and the network becomes the backplane. Ahead of next week's OCP Summit, Arista launched an open rack-scale portfolio led by Etherlink SU-144, an Ethernet scale-up architecture that connects up to 144 accelerators in a single-hop domain and up to 1,024 across racks, built with AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft and Qualcomm. The argument is that in the AI era the network stops being a pipe between servers and becomes the thing compute is built around — though Arista is candid that Nvidia's proprietary stack still owns the near term for its own systems.

Policy

• ICANN's new domain round is an AI land grab. The application list is out: 1,615 requests from 481 applicants, with ten companies — Meta and OpenAI among them — competing for .agent, seven for .agi and six for .asi, while OpenAI applied for 15 different suffixes, Meta for 21, and Anthropic wants .claude and .anthropic. Multiple applicants for the same string land in contention sets, so the fight to brand the AI era's real estate now moves to ICANN's adjudication queue.
• Trump's AI science summit lands Thursday with more than $1 billion in industry commitments. The White House will announce pledges from AMD, OpenAI, Anthropic and others toward the Genesis Mission, plus a compute hub spanning more than 14 institutions across 10 southeastern states and a $1 billion investment in Georgia for student and faculty compute access; the NSF and Energy Department will add more than $100 million for AI-driven science tools. The subtext: with federal science R&D budgets trimmed, industry is being asked to fund the lab bench — and to frame it as super-intelligence leadership.
• The first real-world "adversarial poetry" attack is running a 3,000-server botnet. Lumen's Black Lotus Labs says the PoeLLM malware, attributed to an Italian-speaking actor, hides its command-and-control address inside a poem posted to GitHub — change the poem, and every infected machine finds the new server, with no link or encrypted blob for scanners to flag. Victims are mostly internet-exposed open-source AI infrastructure (LiteLLM, Ollama, Gotenberg, Gitea) across the US and Western Europe, mined for cryptocurrency and conscripted as scanners; at its peak the campaign infected more than 800 servers a day.

Tools

• Docker open-sourced its agent runtime. Docker Agent is a builder and runtime for multi-agent systems: declarative configuration, built-in tools plus any MCP server, support for essentially every model provider, and agents packaged as OCI images you can push to a registry and run anywhere. It ships pre-installed in Docker Desktop, which means "works on my machine" now extends to whole agent teams.
• AWS shipped an open-source sandbox that doesn't take the agent's word for anything. Strands Box combines OS-level isolation with a policy engine that remembers what the agent has already done — so rules can cap API calls, rate-limit posts to Slack, or gate a Git push, enforced deterministically rather than by asking the model nicely. It's on GitHub now, macOS-only for the time being, with Linux support in development.
• llama.cpp got better at running giant mixture-of-experts models on small GPUs. Two changes landed this week: a GPU-side cache for experts that have to live in host memory — sparing the slow trip from system RAM on every pass — and an optimization for GLM5Next's multi-token prediction. Translation: smoother local inference on consumer cards for exactly the class of models that otherwise stutters.

Is validation — not discovery — the real bottleneck in AI security now, and who should pay for fixing what the agents find? Tell us in the comments.