Planet AI Weekly

October 25, 2026

This week we reviewed 927 articles from 48 sources.

PHP core vulnerabilities discovered by AI put 70% of websites at immediate risk. Recursive shipped Flint, a coding model built by its own research system. NVIDIA expanded European sovereign infrastructure with Vera CPUs and open-sourced Run:ai, while Microsoft framed agents as insider threats and the security architecture race accelerated.

Official Highlights

Anthropic donated Agent Skills to the Agentic AI Foundation. The open standard, adopted by 70-plus platforms including Cursor and GitHub Copilot, now lives under vendor-neutral governance. Read more

NVIDIA open-sourced Run:ai under Apache 2.0 for Kubernetes GPU orchestration, with scheduling, quota management, and GPU fractioning now inspectable and self-hostable. Read more

European cloud providers Nebius and Nscale are building local infrastructure on NVIDIA Vera CPUs to power agentic computing and data processing. Read more

German Chancellor Friedrich Merz and NVIDIA CEO Jensen Huang detailed plans to merge German engineering with state-of-the-art AI, from automotive digital twins to autonomous vehicle development. Read more

Ollama added speech synthesis locally via Phonon from Gradium, streaming speech over 100 times faster than real time on Apple Silicon in five languages and dozens of voices. Read more

Editor's note: In regards to Ollama adding speech I am happy to see it and great that they are looking to add more beyond just mac. In the meantime THIS is a great ElevenLabs alternative open source solution: https://github.com/debpalash/VoiceStudio at the time of this post 86k stars on github.

From the Community

PHP2Shell (CVE-2026-XXXXX): RCE in PHP core since 2004. OpenAI's Patch-the-Planet program and Google's Wiz found it independently. Exploits in the wild within hours. Patch now. Read more

Editor's note: On Debian the fix is a package refresh, an upgrade and an Apache restart, about five minutes. The catch: if your server's package list has not been refreshed in a while, it will not even show the fix as available until you run apt-get update. Fixed versions are 8.3.36, 8.4.27 and 8.5.12. If you are on 8.1 or older there is no patch, so upgrade. If you are on shared hosting, ask your host when they patched.

Microsoft treats AI agents as insider threats because frontier models can't be trusted, launching Guardian Agents to monitor what employees deploy alongside Project Perception's red/blue/green team framework. Read more

DeepSeek DeepInfer published the only public breakdown of frontier-scale inference economics: 155T tokens per day per cluster at roughly 80% margin, yielding $22M daily revenue and 81-day hardware payback. Read more

Kimi K3.1 Pro, a 104B A-2.8T model from Moonshot, narrows the closed-AI lead to one month by credible evaluations, with full weights promised and a highspeed API already live. Read more

Editor's note: if you are not already... you can also use kimi inside claude code. Read more and more

Recursive's Flint ships as an open-weight coding model under Apache 2.0, built largely by its Eureka Machine automated research system in under four months. Read more

NVIDIA's OpenShell reference architecture isolates each AI agent in a kernel-level sandbox on Vera CPUs with policy enforcement via BlueField DPUs. Available now as a blueprint, not a product download. Read more

Microsoft open-sourced Agensh, a multi-agent harness scaling to 1,024 agents with no central orchestrator, using shared Git repos and chat for self-organization on ProgramBench. Read more

Google Cloud announced Agent Compute Engine, giving each AI agent a hardware-isolated microVM with zero credentials inside, already powering Gemini agents at work. Read more

Featured This Week

DeepSeek DeepInfer is the first public ledger of frontier inference unit economics. The breakdown — 155T tokens per day per cluster, ~80% margin, $22M daily revenue, 81-day hardware payback — gives infrastructure teams a concrete baseline for modeling GPU capex that previously required guesswork. For a field where every lab keeps its P&L private, these numbers reset how practitioners size training versus inference clusters, negotiate cloud contracts, or justify on-prem builds. The 81-day payback in particular reframes the depreciation math that governs whether to buy or rent. If the figures hold, they also explain why inference specialization is attracting capital faster than training clouds: the cash turns over faster than the silicon depreciates. Read more

Send comments and story tips to tips@planet-ai.net. If you found this useful, share it with a colleague.