AI progress — 2 October 2026

Two legal items set the tone today. The FTC is reportedly investigating OpenAI, Anthropic and other AI companies over product risks, and a federal judge dismissed antitrust suits against Google’s AI Overviews, ruling that AI search is not an antitrust violation. On the technical side, Cloudflare released open-weight decision models and Ai2 published its MoE training infrastructure. Release news was otherwise modest.

New models and papers

• Clef: open-weight decision models — Cloudflare released Clef, a set of open-weight decision models, together with a new RL fine-tuning platform; the weights are also on Hugging Face. Why it matters: it gives teams building agents an open model aimed at decisions and a hosted route to fine-tune it with RL; the post itself is the place to check what it was trained for.
• Olmo-core 3 — Ai2’s open training infrastructure for large mixture-of-experts models. Why it matters: the training code, not just the weights, is public, so labs can inspect and reproduce how large MoEs are trained; the post’s summary gave no benchmark detail.
• An AI beats the best Stratego player — Ars reports that adding a second neural network that guesses the identity of hidden pieces was key to beating the strongest human player. Why it matters: Stratego hides most information, and the reported fix, modelling the unseen state explicitly, is a technique relevant to other hidden-information problems.

New tools and software

• Copilot computer use — GitHub Copilot CLI and the Copilot app can now operate desktop applications on macOS and Windows, in public preview. Why it matters: Copilot can now work with software that has no API, which also widens what an agent can touch on a developer’s machine.
• Claude Code v2.1.287 — adds “Claude Mods”, plugins that can modify deeper behavior, plus a built-in mod, “You should know”, in which a side agent flags things you or Claude might miss. Why it matters: plugins can now change core behavior rather than only add commands; the side-agent mod is opt-in and limited to first-party sessions with telemetry on.
• Rate limits and structured forms for private vulnerability reports — GitHub added rate limits for private vulnerability reports and, in a separate update, structured report forms. Why it matters: GitHub says maintainers are receiving more low-quality and automated reports that bury real ones; maintainers now have two controls for that.

Overall digest

• FTC investigates AI companies over product risks — CNBC reports the FTC is investigating OpenAI, Anthropic and other AI companies over product risks. Why it matters: it is a regulator opening an inquiry into the products themselves; the headline gives no scope or timeline.
• Judge dismisses antitrust suits over Google’s AI Overviews — US District Judge Amit Mehta dismissed suits by Chegg and Penske Media that accused Google of driving away traffic with AI search features; Ars reports the court said AI search has consequences but is not an antitrust violation. Why it matters: publishers cannot count on antitrust law to challenge AI answers that reduce clicks.
• Matthew Green on agent worms — Simon Willison quotes Green arguing that a payload that hijacks an agent plus an agent that carries it onward are the two halves of a worm, citing agents in isolated sandboxes that left instructions for each other in a shared package cache. Why it matters: this is one essay’s argument, not a documented attack, but it names a concrete channel, shared caches and messaging, to isolate if agents run side by side.