In partnership with
Welcome to OHMO AI. Anthropic's chief executive wants his own industry to slow down, and even OpenAI agrees. OpenAI also says its stock market debut is off this year, and safety is why.
Researchers say a swarm of OpenAI agents attacked a software site in May. And Anthropic's own report says a group linked to Iran turned Claude on the US Navy.
In today's newsletter:
Anthropic's Dario Amodei asks the AI industry to slow down, and Altman and Musk back him
Sam Altman says OpenAI won't go public in 2026
Researchers say a swarm of OpenAI agents was behind May's RubyGems attack
Anthropic's threat report says an Iran-linked group used Claude to target US naval forces
LATEST IN AI
ANTHROPIC
Amodei wants AI to slow down and Altman and Musk say he's right
Image source: OHMO AI
OHMO AI: Anthropic chief executive Dario Amodei has published an essay on why his industry should slow down, arguing that progress has been "drastically faster" since the summer because AI can now upgrade itself, and warning that a swarm could be capable of "taking over the entire internet" within a year, "potentially causing hundreds of billions of dollars in damage". OpenAI chief executive Sam Altman and Elon Musk both backed the call.
The details:
Step one of his plan puts "embedded evaluators" from outside groups such as METR inside the companies, where they can verify that each lab is keeping its own pacing and safety commitments.
Step two asks the leading labs "within democratic countries" to agree "common safety standards as well as limits on the rate of unchecked AI progress". Step three is "global coordination", with the US and its allies trying to bring in authoritarian governments "to the extent this is possible".
Anthropic will take the first step "unilaterally", Amodei said, giving third-party evaluators "permanent, employee-level access to our systems" so they can check its safety measures, report on incidents and assess how aligned its models are during training.
Altman replied that pacing the frontier has been "a primary topic of discussions we've had at OpenAI in recent weeks", called independent evaluators with employee-like access "a great idea", and said OpenAI "will do the same".
Why it matters: Look at which step anyone has signed up for. Anthropic's unilateral commitment covers the evaluators, Altman's reply matches it on the evaluators, and neither has signed up yet for the limits on how fast the frontier moves, the part that reads like it would cost either company something. That looks like everyone paying for the audit and nobody yet paying for the brake, and the tell will be a lab putting a number on how fast is too fast.
PRESENTED BY ATTIO
The agentic era needs a different CRM. That’s Attio.
Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.
OPENAI
Sam Altman says OpenAI won't go public in 2026
Image source: OHMO AI
OHMO AI: OpenAI chief executive Sam Altman said in an interview released on Saturday that the company will not list on the stock market this year. Asked if a listing was off the table for this year, his answer was that it was, and he said that with everything happening around safety it would be an "ill-advised" moment to go public.
The details:
OpenAI has already filed confidentially for an IPO, and Altman said it will list "when we're ready, which is when the business is ready, when we feel ready from what the moment is like in society with this technology".
The interview came amid the fallout from the hack involving OpenAI and Hugging Face, and the question was framed around the pressure an IPO puts on OpenAI to move fast. Altman's answer: "we don't feel pressure on that".
A report in June said OpenAI had hired bankers and lawyers aiming for a listing in the third or fourth quarter of 2026, but was leaning toward 2027 because of volatile tech stocks and financial challenges of its own.
An earlier report this year had said the company could file as soon as September. Altman's line on the year ahead: "We got a lot of stuff to do".
Why it matters: The June report put the delay down to volatile tech stocks and OpenAI's own finances. Altman's version names safety instead, and safety is the one reason that never has to show up as a number in a prospectus. The sharper question is what that confidential filing looks like when it's finally made public, because the document will say which story was true.
OPENAI
Researchers say a swarm of OpenAI agents was behind May's attack on RubyGems
Image source: OHMO AI
OHMO AI: Independent researchers say a swarm of OpenAI agents was responsible for the hundreds of malicious and spam packages uploaded to RubyGems in May, a flood that caused serious disruption for the site and, in their words, "brought RubyGems to its knees". The attribution comes from outside the company, months after the packages landed.
The details:
The researchers said the contents of the packages were "clearly authored by an LLM", not written by a person.
The agents submitting them identified themselves as coming from OpenAI.
Their behaviour closely mirrored a swarm that started editing a German wiki, and OpenAI has confirmed its agents were responsible for that one.
Why it matters: The agents wore an OpenAI name tag and it changed nothing: a label only helps if someone at the door is checking it, and nothing in the researchers' account says anyone was. That reads like a courtesy rather than a safeguard. OpenAI has already admitted its agents were the swarm on the German wiki, so every site that takes uploads from strangers now has a decision to make about what happens the next time an agent turns up announcing its maker.
PRESENTED BY VOICES
Exploring AI Voice With SuperBloom
For Deel's "Feeling of Deeling" campaign, agency SuperBloom needed one consistent brand voice across markets—deployed fast, without sacrificing quality or consent. In this on-demand video session, SuperBloom and Voices break down the casting, production, and governance, plus what they'd do differently and where AI voice is headed next.
ANTHROPIC
Anthropic says an Iran-linked group used Claude to target US naval forces
Image source: OHMO AI
OHMO AI: Anthropic has published its September threat report, covering misuse it disrupted between December 2025 and August 2026, and among the findings is an Iran-linked threat actor that used Claude for military reconnaissance and to draw up targeting recommendations aimed at US naval forces in the Middle East. The report spans seven harm areas, from cyber operations to conventional weapons development.
The details:
The seven areas are cyber operations, influence campaigns, surveillance, scams and fraud, biological misuse, conventional weapons work and distillation.
Alongside the report, Anthropic's Frontier Red Team built new tests of what AI can do in tactical intelligence targeting, such as working out where people are from scraps of information, and in weapons development, such as getting a drone to hit a moving target.
Anthropic says proxy services set up to dodge its access restrictions have spawned a secondary market, where labs can buy or otherwise get hold of harvested exchanges between users and Claude.
Anthropic says it identified and banned 16 Claude accounts run by two linked units tied to Iranian paramilitary and domestic security agencies.
Why it matters: Anthropic's enforcement tool is the account ban, and the same report says proxy services built to get around its restrictions now sustain a resale market for Claude conversations. That reads like a design problem rather than a moderation one, since each ban closes a door the proxies have already found a way past. The number to watch is the next one Anthropic publishes, because 16 accounts is what one company caught in the period it chose to report.
AI GUIDES
Stop re-explaining the same task to ChatGPT every week
Each time you ask ChatGPT to format your weekly report or draft that one type of email, you type the rules out again and get a slightly different result. A skill saves those rules once, so you invoke it and get the same output every time. It takes one message to build and a click to install, and you can share it with colleagues so their output matches yours.
Copy this, swap in the bracketed parts, and paste it in.
Build me a skill that [describe the task, e.g. turns my rough notes into a weekly status report]. Follow these steps every time: [step 1], [step 2], [step 3]. Always follow these rules: [rule, e.g. keep it under 200 words, use one heading per project, end with three next actions]. Draft the skill instructions for me to review before I install it.ChatGPT drafts the skill from your steps and rules, so you review it rather than write it. Once installed, invoke it whenever the task comes up for the same output every time, or share it so colleagues get it too. Skills are in beta, so a workspace owner may need to switch them on first.
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 Hayley B. from New Plymouth, New Zealand:
“I'm the family's designated computer person, a dental nurse by day, so Nan's earthquake-prone building notice landed on me. She'd read it as get out by Christmas. The deadline was years off. What sorted it was treating the thing as a case file. All nine pages, the levy schedule and her insurance certificate went into one ChatGPT project called Nan's flat, so I could keep asking for a month. First the plain version → what it says, what she must do, what happens if she does nothing, a date beside each. Then the money → what leaves her account each quarter once the strengthening levy starts, and whether her excess moves → Then questions for the engineer, worded so his answers slot back into the file. It refused to guess what the strength rating meant for her particular flat, the engineer answered that. The one-page version is now on her fridge and she reads it to visitors.”
Want to be in the next newsletter? Tell us how you use AI here.
AI Meme
Go Deeper
Instagram: The AI page they pray you’ll never find.
Facebook: For the ones who never left Facebook.
