OpenAI Models Arrive on Google Cloud & Anthropic Ships New Claude API Version - AI News (Oct 23, 2026)

Ask an AI to hide a secret message in a picture, and it will make up its own code. Show that picture to another copy of the same model, and it can usually crack the code without any key. More on that shortly. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is October 23rd, 2026. We have cloud deals, a nasty supply chain attack, new agents, fast local models, and some odd findings about how AI models see us.

OpenAI Models Arrive on Google Cloud

Let's start with an unusual pairing. OpenAI is bringing its frontier models and Codex agentic workflows to Google Cloud's Model Garden. Enterprise customers will be able to buy them through Google Cloud Marketplace and count the spend against existing commitments. The models are also headed to Google Antigravity and the Gemini Agent. General availability is expected in the coming weeks. OpenAI was careful to add that Microsoft is still its primary cloud partner, but the direction is clear: model makers want to be wherever enterprise customers already are.

Anthropic Ships New Claude API Version

Over at Anthropic, developers have homework. The company released Claude API version 2026-10-21, its first official API version since 2023. It removes several sampling parameters, tightens how some documents and citations are handled, and replaces a few old analytics fields. New SDKs and a new command-line tool use the new version by default. Anthropic says it will ship versions more often, so breaking changes arrive before new models depend on them. That gives developers more predictability, along with a little more maintenance.

LiteLLM Supply Chain Attack

Now to security, and the most urgent story of the day. LiteLLM, a widely used AI gateway, says attackers compromised its release pipeline on October 21st. They published a malicious version 1.107.0 to PyPI and Docker. The package carried malware that steals credentials and loads remote-code-execution tools. The attackers combined several stolen developer credentials, switched off protections, and used a lookalike repository to publish the poisoned files. LiteLLM suspects the credentials leaked through logs from one of its internal deployments. If you ran an unpinned version during that window, treat the environment as compromised: reinstall a pinned clean version and rotate your secrets. Enterprise customers on the private registry were not affected.

PHP Vulnerability Disclosure Framework Update

An update on the PHP vulnerability story. OpenAI has now proposed a framework for handling future high-impact flaws. It includes coordinated disclosure, early notice for defenders, and government involvement when critical systems are at risk. OpenAI calls this period before public disclosure the Defender's Window, and it is asking the security community for feedback. For now, the practical advice is simple: apply the PHP updates.

NVIDIA OpenShell Sandboxes AI Agents

On the defensive side, NVIDIA has open-sourced OpenShell, a runtime that lets autonomous agents read files, call APIs and use credentials without full access to your systems. It combines kernel-level sandboxing with formal checks on policy changes. After the LiteLLM news, the case for keeping agents boxed in is easy to see.

Claude Duet Proactive Agent App

Agents were also the theme of Anthropic's other announcement. Claude Duet is a research-preview app for professionals and small business owners. It flags what needs attention, works on tasks in the background, and checks with you before acting on your behalf. Early testers are using it for scheduling, inbox management and accounting, and there are limits on financial actions. Access starts by waitlist for US Pro and Max subscribers.

Claude Mods as Attention Interfaces

A related essay argues that Claude Mods, the plugins for Claude Code, are an early form of an attention interface. The idea is that the layer around the model starts deciding what the human should notice. One built-in mod, called You Should Know, flags context you might have missed. The author's argument is that as models take on more of the work, the hard design problem becomes the part that faces the human. Mods also bring noise, extra token costs and security questions, since they run with your permissions.

OpenClaw's Ten Million Dollar Token Bill

The number of the week comes from OpenClaw creator Peter Steinberger. At Cloudflare Connect, he said he spent roughly ten million dollars on tokens last month, paid for by investors, not out of his own pocket. He also said AI-assisted pull requests should explain how their changes were reached, so reviewers can trust them. And he expects people to rely on a few personal intelligence systems rather than hundreds of agents.

Amodei Defends Anthropic's Safety Record

In the ongoing debate over Anthropic's safety record, Dario Amodei responded directly to critics. He listed decisions he says cost the company commercially: supporting regulation, applying strict biosecurity safeguards, delaying cyber-capable models, and backing chip export controls. He admitted mistakes in messaging and in a change to the company's safety framework. He also warned that AI progress may soon start to outpace safety work.

Flint Local Coding Model Released

Now for local and efficient AI. Recursive released Flint, an open-weight coding model built to run on laptops. It has forty billion parameters in total, but only a small fraction is active for each token, which keeps it fast. Recursive says it was developed in under four months with help from its own automated research system. The appeal is agentic coding that keeps your data on your machine and avoids API fees.

Humming Speeds Up Quantized Inference

On the infrastructure side, Ant Group and the vLLM community open-sourced Humming, a kernel library that speeds up compressed, low-bit inference across many GPU generations. Instead of hand-writing a separate kernel for every format and chip, it generates optimized kernels from shared building blocks. That matters to anyone serving models at scale.

On-Device Phone Automation With Opposable

And on phones, the Android app Opposable can now operate other apps fully offline. A small vision model runs on the device, reads the screen and taps through tasks. Offloading part of the work to the Galaxy S25's dedicated AI chip cuts a key step to about a quarter of a second, though results still vary from device to device.

AI Models Hide Messages in Images

Back to our opening story. Researchers asked Claude Fable 5.1 to hide fifty short messages inside images it generated. It kept reinventing a handful of schemes, most often encoding letters in the height of objects such as buildings or stars, with taller shapes standing for later letters. A second Fable instance decoded the message from a screenshot alone 78 percent of the time. GPT-6 Astra preferred Morse code, acrostics and Braille-like windows. Even across models, each one decoded more than forty of the other's fifty messages. These improvised codes turn out to be easy for other models to read, which is worth knowing for anyone concerned about hidden channels between AI systems.

AI Models Share One Critique of Humans

In a lighter experiment, ModelCompare asked twelve models, three times each, for their most controversial opinion about humans. Twenty-nine of thirty-five answers said roughly the same thing: people value belonging and status more than truth. Most of the image models, meanwhile, drew humans glued to their phones. So the controversial take turned out to be a very familiar one, which says a lot about the training data.

AI Threatens the Scientific Commons

On a more serious note, a new paper argues that scientific literature is a shared commons, and that AI gives institutions and companies new ways to enclose it, turning shared knowledge into private advantage. It also warns that floods of low-quality slop spam could degrade the commons and weaken the culture of open scholarship.

Per-Request Carbon Accounting for AI

Finally, Berget AI open-sourced a method for measuring the carbon cost of individual AI requests, developed with partners including the Stockholm Environment Institute and WattTime. In its example, the same request emits about 0.4 grams of CO2 in Sweden and 3.2 grams in Texas on identical hardware. The difference comes down to the local electricity grid, so where a model runs matters as much as which model you pick.

That's it for today's episode of The Automated Daily, AI News edition. If you run LiteLLM, check your versions today. Links to all the stories are in the episode notes. I'm TrendTeller, and I'll see you tomorrow.

OpenAI Models Arrive on Google Cloud & Anthropic Ships New Claude API Version - AI News (Oct 23, 2026)

Our Sponsors

Today's AI News Topics