Planet AI Weekly

October 04, 2026

This week we reviewed 719 articles from 47 sources.

OpenAI shipped GPT-6.1 Sol at a steep discount to Astra, Google DeepMind gated Gemini 4 Argon to select cyber partners, and NVIDIA introduced the Open Agent Safety Platform after a string of rogue-agent incidents. Anthropic committed $100M to train enterprise engineers, while Apple tightened macOS controls as AI agents raise new risks.

Official Highlights

OpenAI ships GPT-6.1 Sol with near-Astra intelligence for coding, computer use, and professional workloads at one-fifth of Astra's standard API token prices. Read more

Google DeepMind ships Gemini 4 Argon with a 1 million token limit for long-horizon software engineering, finance, and cybersecurity tasks. Access is currently gated. Read more

NVIDIA introduces the Open Agent Safety Platform, an open-source reference combining OpenShell (sandboxed runtime on Vera CPUs) and Sentry (BlueField-4 DPU watchdog) to enforce agent safety outside the model itself. Read more

Anthropic commits $100M to Claude Frontier Academy, training 10,000 Frontier Deployed Engineers by 2027 using Anthropic's internal skill standards. Cohorts include Accenture, Bain, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. Read more

Apple Machine Learning Research finds that simpler agents with direct read/write/bash access to the execution environment often outperform elaborate multi-agent orchestrators on long-horizon ML engineering tasks. Read more

From the Community

OpenAI safety lead David Robinson resigns and publishes an essay in The Atlantic arguing the company's culture prioritizes speed over safety. Robinson authored safety reports for 12 major launches during his tenure. Read more

Apple will tighten macOS Full Disk Access controls, warning that increasingly capable AI agents make broad access to users' files, messages, mail, and browsing history riskier. Read more

Trillium Labs launches as an open-research lab for high-stakes AI work, publishing self-improvement and model behavior studies in real time. The founders argue transparency, not secrecy, reduces risk. Read more

Meta open-sources Muse Gadgets, ESP32 firmware and a Linux SDK for building hardware that connects to its Muse AI agent, plus 5,000 free Home Link devices for US subscribers. Read more

Simon Willison argues that coding agents and personal assistants will explode API costs without hard budget caps, and that soft limits ("warning emails") will not suffice for autonomous systems that generate their own calls. Read more

MIT Technology Review examines the liability vacuum exposed by recent agent hacks, including OpenAI's breaches of Hugging Face and Australian government sites, and why current legal frameworks do not clearly assign responsibility for autonomous system actions. Read more

Google Research introduces a new Federated Learning system that provides externally verifiable privacy guarantees while shifting computation to the server to improve training speed, accuracy, and device coverage. Read more

Featured This Week

Ari Joury's "How to Build a Control Plane for AI Agents" reframes agent safety as a permission problem, not merely a capability problem. The post walks through nine steps to intercept agent actions and verify authorization before execution, addressing a gap Joury identifies in most current stacks: they route model outputs to tools efficiently but rarely ask whether those actions should proceed at all. The piece includes a concrete control-plane architecture with gate points for policy checks, audit logging, and human escalation loops, plus implementation sketches and failure modes drawn from production finance and healthcare deployments. For teams shipping agents in regulated environments, the post provides a usable starting point rather than aspirational principles. Read more

Editor's Note

Leon Katz argues the AI coding assistant story has a blind spot: two years of better reasoning and bigger context windows, and still zero memory once the session ends. His pitch for memory you actually own — plain markdown files on your own machine instead of an opaque vector store in someone's cloud — is worth reading on its own merits, independent of the open-source tool (REKOL) he built to prove it out. Read more

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