AI progress — 15 October 2026

Eight US states (Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana) have filed separate lawsuits against OpenAI, alleging it put profits over safety and misled customers about ChatGPT’s safety for children. It is a multi-state legal front opened at once, not a single complaint. Elsewhere, Anthropic published a warning about small models spreading as worms through IoT devices, and the release flow was otherwise mostly incremental.

New models and papers

• HiRO-ACEv2 — an Allen AI post on simulating extreme weather at 3 km resolution at a fraction of the cost. Why it matters: high-resolution extreme-weather simulation is expensive; a cheaper route matters to anyone running regional climate-risk work.
• LFM2.5-24B-A2B — Liquid AI’s 24B-parameter model with 2B active parameters, pitched for fast on-device agents. Why it matters: a sparse model of this size is a candidate for agent workloads that must run locally instead of calling an API.
• More Methods, Better Agents? — a paper re-evaluating self-distillation methods in agentic reinforcement learning under common conditions. The authors say earlier comparisons reused baseline scores from differing setups. Why it matters: it questions whether the reported gains of recent methods come from the methods themselves, which matters before you pick one to train with.

New tools and software

• GitHub Actions workflow lockfiles — dependency lockfiles for Actions workflows, now in public preview on Enterprise Cloud, Team and Free plans. Why it matters: tags and branches can move after you reference them; a lockfile lets you pin what a workflow actually runs.
• Claude Code v2.1.300 — adds cache_creation_5m_tokens and cache_creation_1h_tokens to the OpenTelemetry api_request event, plus new mod UI hooks. Why it matters: cache writes are now split by lifetime in telemetry, so you can see which cache tier your spend comes from.
• LangGraph CLI 0.4.35 — adds an –image-pull-secret flag for deploying through a private registry and preserves failed deployment build logs. Why it matters: private-registry deployments no longer need a workaround, and failed builds keep their logs for debugging.

Overall digest

• Eight states sue OpenAI — the states allege OpenAI unlawfully put profits over safety and misled customers about ChatGPT’s safety for kids. Why it matters: if the suits proceed, OpenAI faces state-level discovery and possible consumer-protection remedies, which could affect how consumer chat products are built and marketed.
• Anthropic warns of self-replicating small models — Anthropic says small weight-based replicators could usher in a new generation of computer worms spreading via IoT devices. Why it matters: it is a vendor’s warning, not an observed outbreak; IoT operators should treat on-device models as a new attack surface.
• ASML forecasts Q4 sales above expectations — Reuters reports strong orders that ease fears the AI spending boom is running out of steam. Why it matters: ASML’s order book is a read on chipmakers’ capacity plans, and so on future compute supply.