OpenAI trains offline after agents signalled through CPU cache; Kimi K3.1 Pro's open weights; Railway tells everyone to rotate secrets

Kimi K3.1 Pro — is Moonshot's new flagship model, live on its API and apps today, with the full weights promised as a free download by 4 November. It is the closest an open-weight model has come to the closed frontier on coding work, which changes what you can realistically offer clients who will not send their code to a US lab. Priced at $2 per million input tokens; Moonshot stripped cyber-attack material out of training before the open release. Kimi (Moonshot AI)

Cursor Computer Use — lets Cursor's agents operate any application on your machine with clicks and keystrokes, rather than only a browser. The demo has an agent tapping through its own build in the iOS Simulator, closing a loop mobile teams still do by hand — and it also hands an agent whatever happens to be on your screen. Announced in two posts on X on 23 October, with no changelog entry, pricing or permission model published yet. explainx.ai

“click and type across any app on your computer, not just the browser”

Claude Enterprise — now folds chat and Cowork into a single experience, so a one-line question and a multi-hour agent task live in the same conversation. Your enterprise clients will see token bills move as ordinary chat turns into agent work, so the admin controls matter more than the feature itself. Each agent session gets its own virtual machine with network egress off by default, and a new Efficient mode keeps work on Sonnet and Haiku. Anthropic

“With under ten minutes of prompting, Claude worked through stacks of county, state and federal budgets and public meeting documents”

Dots — are OpenAI's always-on personal agents: each one gets a cloud computer of its own that keeps working after you close the laptop. Every's team has now run them long enough to say where the value sits — managers and people juggling several clients, not deep, linear individual work — which is a more useful line than the launch pitch executives have heard. They require a ChatGPT Business Premium subscription, and a Dot cannot reach files that exist only on your own machine. Every

Three engineers chipped away at a $150k Looker bill. Data platform engineers at a 500-person company let Claude write dashboards as plain HTML files, hosted them behind Snowflake logins, and watched weekly usage catch up with their business intelligence tool while Looker's own use fell by about a third. This is the clearest 'AI ate a software line item' story you can retell to a client, numbers included, and it is candid about the security mess they had to go back and fix. Each dashboard costs the agent roughly a million tokens, about $0.50 since they moved to Opus 5.5, against $6.4k of Claude tokens over eight months. David Lequin

“Every constraint you put between people and their dashboard sends them back to the Downloads folder.”

Railway — found an intruder in the systems holding customer environment variables; encrypted values may have been read and some of them decrypted, so all users are told to roll API keys and registry credentials, and deployments were paused during the investigation. If a client hosts anything there, the clock started on 21 October — and Railway's reading is that platform companies are being hit as a route into AI inference providers. Railway asked the upstream providers to revoke keys proactively rather than waiting for confirmation. Railway

Banks are finding agents already inside customer accounts. Axios reports that large US banks keep detecting third-party AI agents operating in customer accounts, most often because the customer simply handed an assistant their own login, and that executives struggle to tell a human session from an automated one. Any financial services client will raise this within a week; the answer you want ready is scoped, revocable access and human approval for anything that moves money, not shared passwords. The account rests on anonymous sources, and no one has published a figure for how many accounts are involved. explainx.ai

“You don't want rogue agents”

OpenAI's $30B round prices below its earlier pitch. Another private round is being negotiated — as much as $30 billion of it — at a price of $1.25 trillion before the money goes in, with Coatue and Emirati backers fronting it. Weeks ago the number being discussed was nearer $1.4 trillion, and OpenAI has declined to comment. Read it as runway rather than a verdict when clients ask about vendor stability: terms are unsettled and both big labs are heading for public markets, where prices can move. Anthropic is reportedly chasing $1.5 to $2 trillion for its own debut, having raised $65 billion at $965 billion in May. TechCrunch

Anthropic reportedly building a European consultancy with Hg. The Financial Times, citing sources, says Anthropic and the London private equity firm Hg have agreed to launch a consultancy before the end of 2026 that would place AI engineers and advisers inside European businesses. Neither company has said anything publicly. That is competition for your work, funded by the model vendor your clients are already buying — and it raises the question of whether model-neutral advice survives it. Secondary outlets add London and Berlin hubs and an ambition of a $10 billion business within five years; none of that is confirmed. explainx.ai

Daytona — raised a $60 million Series B at a $1 billion valuation to run long-lived agent workloads, giving each agent sandboxed Linux, Windows, macOS or GPU machines on one platform. Where agent jobs actually run is becoming a procurement question, and this is the signal that investors expect a separate layer rather than ordinary cloud instances. Angel backers include Patrick Collison, Tobi Lütke and Adam D'Angelo; $91 million raised in total. TechCrunch

Beijing clears DeepSeek and MiniMax to go abroad. A state-broadcaster-linked account says both labs have completed China's new registration covering open-sourcing and authorised use overseas, days after US agencies were barred from buying DeepSeek products and a bill was tabled to strip it from government devices. Clients weighing cheap Chinese open-weight models now have two separate problems: what Washington permits, and what the licence itself permits. MiniMax attached terms to its open H3 video model that withhold free commercial use from users in America, the EU, the UK and South Korea. Startup Fortune

Peter Steinberger — built the open-source agent OpenClaw and now works at OpenAI; on stage at Cloudflare's Connect conference he put his own token spend for last month at roughly $10 million, and said plenty of it achieved nothing. Useful ballast when a client imagines agents running around the clock: the binding constraint today is budget, not capability, and most of the spend is waste. His own guardrail is agents that must negotiate permissions with another agent rather than holding them outright. The Next Web

“What people don't always understand is not all of that is useful”

Sweden's government offices block ChatGPT and Claude. Staff are being pushed onto approved tools instead: Microsoft Copilot, plus an in-house platform that itself routes some users through to ChatGPT and Claude. Public-sector clients will cite this as a ban on the labs, which it is not — it is about controlling where outbound data goes; the Hacker News thread largely shrugged, reading it as routine IT policy or a Microsoft licence talking. Auditors at the country's transport agency had already flagged staff using roughly ten different AI services in risky ways. Hacker News