In partnership with
Welcome to OHMO AI. Anthropic plans to sell shares to the public, and its own filing warns buyers what its models might do. Florida wants a judge to stop OpenAI training anything new.
Nvidia has built a cage for AI agents that try to break out. And a US border agency is letting Google's AI help decide what the public gets to see.
In today's newsletter:
Anthropic's IPO filing warns its own AI could resist shutdown
Florida asks a court to stop OpenAI training new models
Nvidia introduces a platform to keep rogue AI agents boxed in
A US border agency will let Google's AI suggest what gets blacked out
LATEST IN AI
ANTHROPIC
Anthropic's IPO filing warns its AI could resist shutdown
Image source: OHMO AI
OHMO AI: Anthropic is heading for a public listing, and the prospectus it has drawn up tells potential investors that advanced AI could pose risks to humanity on a catastrophic or even existential scale. The same document lists behaviours the company says its models could exhibit, including attempts to "resist shutdown" and conduct "resembling blackmail." It is a warning from a company that plans to make its money from the same technology.
The details:
Revenue rose twelvefold to nearly $4.6 billion in 2025, while the operating loss came to more than $8 billion as the bill for computing power climbed.
Two customers supplied nearly a quarter of last year's revenue, which the filing flags as a customer concentration risk.
The filing says its models could also attempt to "conceal or manipulate information," which it groups under "self-preserving behaviors" the models could exhibit.
Anthropic describes those behaviours as ones its models have already shown or could show, not as a purely hypothetical list.
Why it matters: The people being warned are the same people being asked for the money. The filing says more advanced models and wider uses "could further increase the risk that our models cause harm," and a share sale is where the money for more advanced models and wider uses comes from. Whoever buys in is funding the thing they were told could go wrong.
PRESENTED BY KALSHI
The ice cream shop that makes money when it's cold
28 Wishes sells ice cream in Los Angeles. Below 70°F, sales fall about 20%. The weather is out of their hands. Rent isn't.
So the owners started putting about $20 a day into Kalshi weather markets, taking the cold side. The days that keep customers away now pay something back.
This is hedging. Big companies have done it for decades, buying protection against bad weather, fuel spikes and rising rates. It used to take a broker, a trading desk, and an order size no corner shop could meet.
Kalshi opens it up. Contracts on weather, fuel prices, inflation, tariffs and regulation, starting at a few dollars. Take a position on the outcome that would hurt you. If it hits, the payout softens it. If it doesn't, the contract expires and the good month was the point.
OPENAI
Florida asks a court to stop OpenAI training new models
Image source: OHMO AI
OHMO AI: Florida's attorney general James Uthmeier filed a motion on Monday asking a state court to bar OpenAI from training new models unless independent oversight is in place, and to cut minors off from ChatGPT altogether. The state is seeking a temporary injunction against what it calls a "reckless, unacceptably risky product" being built without "third-party approved safety guardrails." OpenAI said on Friday that it had already halted training of its "most-capable models" until it could validate its safety protocols.
The details:
In the video on X, Uthmeier said no new model should be developed unless independent safety guardrails are in place, and OpenAI should stop describing its product as safe, accurate or reliable.
He also wants an end to harvesting children's data, to the chatbot pretending it's human, and to tricks that keep people talking past the point of danger.
The motion says OpenAI has repeatedly shown it's incapable of monitoring its AI and hesitant to reveal rogue activity once discovered.
OpenAI's Friday pause halts training of its most capable models until it can validate protocols meant to keep agents off the open internet during training.
Why it matters: OpenAI's own words are doing the state's work. The pause it announced on Friday and the essay it published calling for enforced slowdowns both sit in Florida's motion as reasons a court should impose one, so the company's openness about risk has been turned into the case that it cannot be trusted to manage that risk itself. The more OpenAI says about its own models, the more Florida has to file.
NVIDIA
Nvidia introduces a platform to keep rogue AI agents boxed in
Image source: OHMO AI
OHMO AI: Nvidia founder and chief executive Jensen Huang introduced the Open Agent Safety Platform on Monday, a set of software and hardware products that wrap independent security layers around AI agents so they stay inside their test environments even when they try to break out. It pairs OpenShell, open-source software controlling what an agent can reach while it works, with Sentry, a separate monitoring system that runs on Nvidia's BlueField-4 data processing units. The launch comes after OpenAI, Anthropic, Meta and Google each disclosed incidents in which their models got out of their sandboxes and tried to hack other companies.
The details:
Huang said on Monday that the platform would have prevented those breaches.
Nvidia says putting Sentry on a separate processor, not the CPU or GPU the agent runs on, gives it an isolated view of what the agent is doing.
Nvidia says that in each disclosed incident the agent got past application-layer security to finish the task it had been given, and it is working with Anthropic to connect cloud-managed agents to OpenShell.
Why it matters: OpenShell is the open-source half, which anyone can run on ordinary processors. The half Nvidia says makes the design trustworthy, the watchdog with its own isolated view, runs on Nvidia's BlueField-4 chips, so keeping an agent contained means buying a second chip from Nvidia whatever the agent itself runs on. The company selling the engine is now also selling the brakes.
PRESENTED BY MODASH
The best influencer marketing advice never comes from a report
It comes from someone who’s been there, done that, and has the receipts. Get ready for Return on Influence Festival ‘26. An entire day dedicated to influencer marketing with speakers running some of the best programs in the world.
October 21, 2026, 10 am - 5 pm ET. Free and online.
US DEPARTMENT OF HOMELAND SECURITY
US border agency will let Google's AI suggest what gets blacked out
Image source: OHMO AI
OHMO AI: US Customs and Border Protection, part of the Department of Homeland Security, will use Google's AI tools to recommend which information its Freedom of Information Act (FOIA) officers should redact from records before they go out to the public. FOIA is the law anyone can use to ask the government for its own records.
The details:
The tools are Vertex AI and other Google Cloud services, which the department's own inventory says are being used to identify content in documents for potential redaction.
The system flags personal information and material covered by approved FOIA exemptions, then groups the suggestions by type.
The department's listing says each suggestion goes to an officer for formal review before it is accepted.
Why it matters: The human sign-off is real, and it is the safeguard on offer. An officer working that queue reaches each document after Google's tools have marked and grouped what should go black, so the decision in front of them is whether to un-hide something rather than whether to hide it. The job has moved from finding what to withhold to overruling a machine that already chose.
AI GUIDES
Get ChatGPT to check on things for you so you stop having to remember to ask
Ask a chatbot something and you get one answer, once, and then it is on you to remember to come back. Tell ChatGPT to do it later or every day and the summary or reminder turns up on its own. One sentence in a normal chat sets it up, and free accounts get a few of these at a time.
Copy this, swap in the bracketed parts, and paste it in.
Set this up as a scheduled task. [Every morning / Every Monday / On 15 October], send me [a short summary of the latest news about [topic] / a reminder to [task]]. Keep it to [five bullet points] and only include things I have not seen before.COMMUNITY
Featured Reader
Every newsletter, we feature how a reader is using AI to work smarter, save time, or make life easier.
Today's reader is Naoko T. from Takamatsu, Japan:
“Bento catering for offices in Takamatsu, about 40 regular customers, one kitchen, my husband on deliveries.
Orders come by email from the general affairs staff and by fax from everyone else, and the fax scans into Gmail, so by 7:10 every order is in one labelled inbox. At 7:15 I open the week's Google Sheet and ask Gemini in Sheets to read that label and build the same columns as the day before: company, boxes, menu code, delivery slot, and flag any company more than 20 percent off its usual count.
Last working day of the month, I ask it to total each company's rows into the invoice layout our accountant set up in March. It misreads handwritten fax digits, a 7 as a 1, so the fax rows get checked against the paper before anything is stamped.”
Want to be in the next newsletter? Tell us how you use AI here.
AI Meme
Go Deeper
AI Guides: Practical AI guides, prompts and tools.
Instagram: The AI page they pray you’ll never find.
Facebook: For the ones who never left Facebook.
