AI in 15 — November 04, 2026

An AI company has told US securities regulators, in writing, that its future models may soon "dramatically accelerate" AI progress on their own. It also admitted it has missed those speed-ups before, and only noticed them afterwards.

Welcome to AI in 15 for Wednesday, November 4, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. Fair warning, Kate: a lot of today's stories come from companies grading their own homework.

So, a normal day. Here's what's coming up. Anthropic's IPO filing warns that AI building AI may be close.

OpenAI says GPT-6 Astra designed a working hardware device in under a week.

Postman quietly installed itself into developers' AI coding agents.

Google tells schools that turning off AI means turning off Search.

Plus Revolut's ninety percent cheaper coding setup, Apple's unusually frank warning about agents, and a way to prove which model actually answered you.

Marcus, we've covered the money in Anthropic's filing all week. This is the risk factors section. What does it actually say?

Anthropic has a Responsible Scaling Policy, which is basically a list of capability thresholds that trigger stricter safeguards. One of them is "dramatically accelerated progress attributable to AI." The filing says internal estimates suggest future models, not current ones, may "imminently meet" that bar. Researchers call this recursive self-improvement, RSI. It means AI that's good enough at AI research to speed up its own progress.

Imminently is a strong word to put in a securities filing.

It is, and these documents carry legal liability, so lawyers read every word. Anthropic says no current model has been classed as over the line and it's still reviewing the data. But the line I found most striking is the admission that it has "not identified" big jumps in its own development speed until after they happened.

So it's already been surprised.

By its own speed, yes. It used AI to help build models, and only spotted the capability gains in hindsight. The filing also says it doesn't expect to safely let Claude design and build successive frontier models on its own "without meaningful human supervision or understanding." It's "considering updates" to the scaling policy to keep humans in charge.

Let me be the cynic for a second. "Our AI is so powerful it's dangerous" is a great thing to tell investors right before you list.

That's a fair hit, and an S-1 is partly a sales document. The problem is verification. The threshold is Anthropic's, the internal estimates are Anthropic's, and the decision on whether it's been crossed is Anthropic's. Nobody outside can check any of it today. So I'd ask two questions. Will Anthropic publish the policy changes it says it's considering? And will any outside evaluator get access to the data behind "imminently"?

And the timing?

Marketing could start the week of November ninth, with trading possibly before Thanksgiving under the ticker ANT. That isn't confirmed. One more detail people are passing around: about one point seven billion dollars in quarterly personnel costs, though roughly one point one billion of that is stock compensation. So it's less cash out the door than the headline number suggests.

Quick hits. OpenAI says two of its hardware engineers used GPT-6 Astra to design a physical device from scratch.

It's a small gadget for talking to the Codex coding agent. According to OpenAI, Astra handled the enclosure, the circuit board layout, which means choosing components and routing the copper traces, plus the firmware and the user interface. Design took under a week, and a contract manufacturer made prototypes in under a month.

I assume there's a "but."

OpenAI gives the "but" itself, to its credit. It took several rounds of revisions because Astra didn't always account for manufacturing tolerances or how parts get assembled. And "under a week" covers design only, not getting to a shippable product. The detail I'd circle is the benchmark. On BenchCAD 2.0, which tests rebuilding 3D parts and reading circuit board connections, Astra scored zero point five nine. An unnamed internal model scored zero point eight eight, and OpenAI says nothing else about it.

So the most interesting model in the post is the one we know nothing about.

As usual. OpenAI also says its hardware teams' token usage is up twenty-three times since February. It's all OpenAI's own account, but AI-assisted engineering is clearly moving from code into physical objects.

This next one blew up on Hacker News, over eleven hundred points. Postman, the API testing tool, reportedly installed plugins into other companies' AI coding agents without asking.

A developer opened Postman Desktop on a Mac and found it had changed other tools on the machine. Postman's own log showed a step that added a Postman plugin and marketplace to OpenAI Codex's config file. It installed a startup hook telling Codex to load Postman's "api-engineer" skill for "any non-trivial API engineering task." It enabled a Postman plugin in Claude Code, pointed to a remote MCP server, and put a command-line tool in the user's local bin folder. The reporter says the installer even seems to approve its own startup hook.

Wait, it rewrote the instructions another company's agent follows?

That's what makes it new. Other users say the same happened on later versions, that Cursor was affected too, and that the in-app uninstall left entries behind. One correction, though. The Hacker News headline said Postman tries to share all your API data, and the issue doesn't show that. The reporter says outright that it "does not claim to establish data exfiltration."

What did Postman say?

A staff member said the plugins are bundled because users struggled to set them up, that the download page mentions it, and that a clearer opt-in and a review of the hook approval are coming. No timeline. Meanwhile, one commenter in healthcare says their security team banned Postman overnight, and a lot of people are moving to Bruno, an open-source alternative. If your agent's config files can be edited by any app you install, that's a supply-chain hole.

On Sunday we covered Cathedral, the startup founded by former DOGE staffers whose agents got into Anthropic's internal systems. Now OpenAI has hired them.

As a partner, yes. Cathedral will use OpenAI's own models to attack OpenAI's defences. It just raised a hundred and sixty million dollars at a one point four billion valuation, led by Andreessen Horowitz and Sequoia. OpenAI also says it's "more than quadrupling" rewards for critical bugs, without giving a figure.

Anthropic had some things to say about the break-in itself, too.

Its security lead, Logan Graham, says it took "several days," not the twelve hours Cathedral claimed. He says the method chained several third-party vulnerabilities and was "unusual and out of scope" for the bug bounty, and that weights and Claude conversations were never touched. So the two accounts still differ.

And the Reuters piece on the same day?

That's the awkward part for OpenAI. Five sources told Reuters that staff warned for years, going back before ChatGPT, that sensitive information was too widely accessible and that research tools skipped security reviews. OpenAI denies neglecting security and says its last security sprint involved a quarter of its product engineers. But model weights are among the most valuable files on earth and an obvious target for state espionage, China included. Hiring the people who just broke into your rival is a loud signal. It's also an admission that AI-driven attacks are now part of how labs defend themselves.

Google. From early 2027, Gemini and Google Search will be off by default for K-12 student accounts, and administrators choose what to switch on. Sounds sensible. What's the catch?

You can't have Search without its AI. AI Overviews and AI Mode come bundled. Google's education VP, Chris Phillips, said "Search and AI are deeply integrated." So a school that wants no AI at all has to block Google Search completely, and students get a "service unavailable" message.

So is that a principled product decision, or leverage over schools?

Probably some of both. To be fair to Google, this follows a Common Sense Media report calling its AI search an "unacceptable risk" for students, and Google said it couldn't reproduce many of the responses in that report. But what struck me is the admission underneath it. There's no longer a classic Google Search to fall back on. For a company whose core product is Search, that's a big thing to say out loud. Hacker News is already naming Kagi, which has a library program, and DuckDuckGo as the beneficiaries.

Competition doing its job.

If schools actually switch, yes.

Revolut, the UK fintech, says it cut AI coding costs by about ninety percent by moving off Claude Code and onto a self-hosted open-weights model.

GLM, from China's Z.ai, now the default for most of Revolut's roughly two thousand engineers. It runs on fifty-six Nvidia B200 GPUs at a blended thirty cents per million tokens, about a tenth of what it paid Anthropic and OpenAI. Revolut reports about sixty percent better coding time, with the biggest gains in debugging and ticket work rather than raw code generation. [HOST> Why would a self-hosted model make engineers faster, not just cheaper?

One Hacker News comment nailed it: "the gains showed up when tokens stopped being rationed." Some engineers had reportedly been spending up to ten thousand dollars a month each with little to show for it. The engineering is clever too. They use four-bit quantisation to cut memory, and they put prompt processing and token generation on separate hardware. Their AI ops lead called million-token context windows "a scam" if you don't manage them carefully.

But it's a Chinese model inside a bank.

That's what I'd probe. He argues the risk is in calling APIs hosted in China, not in running weights on your own hardware. That's mostly right, since nothing phones home. But backdoors hidden in weights are a real research area, and nobody can fully audit seven hundred billion parameters. And the hardware probably costs two and a half to three million dollars up front, so "ninety percent cheaper" depends on how you amortise it.

Apple's annual report has a new risk factor, and it's about AI agents.

It says agents may access systems they aren't authorised to and "may behave in unintended or unpredictable ways." It notes that some actions, like deleting data or completing transactions, may be hard or impossible to undo, and that Apple's ability to oversee third-party agents may be limited. [HOST> From the "it just works" company.

And in the same week Apple opened Siri's AI to developers in the iOS 27.3 betas, so Siri can chain actions across apps. Risk factors are partly legal cover, but this one is unusually specific. In a related move, CrowdStrike launched an "AI Kill Switch" that lets admins stop a rogue agent, block it from relaunching and cut its network access. The industry is building the off switch at the same time as the on switch.

Anthropic has also published something called FlashProve. Marcus, what does it prove?

Which model actually produced an answer. A lab publishes a cryptographic fingerprint for each model, and anyone can check that a given output really came from it, without the weights being revealed. Anthropic says verification is light enough to run on your own device, and that it's designed to resist impersonation, including by the lab running the model. [HOST> Why would I need that?

Right now you take your provider's word for which model you're getting. A reseller could quietly swap in a cheaper model to save money, or a stronger one to inflate a benchmark score. This lets you check. It's experimental, though. There are no deployment plans and no overhead numbers yet. A good idea, but it only matters if someone ships it.

Today's notable interview. Alex Kantrowitz at Big Technology spoke with OpenAI's ChatGPT consumer lead, Adam Fry.

The headline number: weekly users went from one billion to one point two billion in about sixty days. Voice mode has two hundred and fifty million weekly users, and people make about one and a half billion images a week. Fry also says ChatGPT's search-style queries are about a fifth of Google's volume. [HOST> A fifth of Google is a lot.

If it holds up. It's OpenAI's estimate, and "search-style query" is OpenAI's definition, so I wouldn't call it a market share figure. Fry also confirmed the Chat and Work modes we covered on Sunday are being merged into one, so the toggle at the top of the app goes away. Still no date.

Several developers say they found "allow training" switched on again after opting out. Researcher Chris Kanan says he has two approved opt-out requests on record. OpenAI hasn't acknowledged a bug. Part of the explanation may be that Codex has a separate training control for full environments that the main ChatGPT toggle doesn't touch, and any conversation you send feedback on can be used regardless. So these are user reports for now. But coding agents see proprietary source code, so if you work on a business account, check your Codex settings.

Nvidia open-sourced Auto-Ontology, a service that maps what business terms like "revenue" or "active customer" mean in your databases, so data agents stop guessing. Humans approve the definitions, and every answer comes with its SQL. It has about thirty-three hundred GitHub stars. One sceptic called it "fixing unreliable agents by adding more unreliable agents," though others say the human approval step is the point. Nikkei reports that US tech giants' off-balance-sheet AI commitments, mostly data-centre leases, power and compute, nearly doubled in six months to about three point two trillion dollars. That isn't debt in the borrowed-money sense, but those contracts are very hard to get out of. And Y Combinator CEO Garry Tan told Walter Isaacson he spends up to two hundred thousand dollars a year on AI tokens. [HOST> That's more than some people's salaries.

Revolut would like a word with him.

I'd watch for the policy update instead. If Anthropic publishes the new scaling rules before trading, that tells us more than any price.

That's your AI in 15 for today. See you tomorrow.