Why OpenAI Pulled a Model That Wouldn't Ask Permission

Dr. Ryan Ries here. This week had a little of everything. There was a major chip deal, a model that got pulled before launch, AI agents flooding restaurant booking systems, and a wild threat report.

Before I start, here are a couple of upcoming events you should check out:

• Charlotte, NC Executive Briefing: Agentic AI for the Enterprise on October 14th
• CPU vs. GPU for Agentic AI: How to get your compute mix right on October 15th
• The Spookiness of AI Token Costs: A Tokenomics Survival Guide on October 29th

Click the links to register, and I may see you there!

AMD Buys World Labs

First, congrats to the AMD team. AMD announced a definitive agreement to acquire World Labs, the AI model and research lab led by Dr. Fei-Fei Li. The all-stock deal is valued at roughly $8.2 billion and should close by the end of 2026, pending regulatory approvals.

World Labs builds spatial intelligence models. These models generate, reconstruct, and simulate interactive 3D environments from text, image, and video inputs. The company works on robotic learning and simulation too. After the deal closes, Dr. Li joins AMD as executive vice president and chief scientist, reporting to Dr. Lisa Su.

Thinking about the logic here... A hardware company buying a model lab is a roadmap move. AMD says World Labs will give it a clearer view of how workloads change as AI moves into reasoning, robotics, simulation, and physical AI. Think of an engine manufacturer buying a racing team. The fastest way to learn where the engine strains is to sit in the pit and watch it run at full throttle.

This hits close to home for me. I spent three years building AR hardware at DAQRI, and getting machines to understand 3D space was the hard problem every day. Seeing that kind of expertise fold into a compute roadmap makes a lot of sense to me.

OpenAI Benches GPT-6.1 Astra

Some bad news for OpenAI. The company confirmed it will not release GPT-6.1 Astra after the model fell short of its safety and alignment requirements. The model was set to launch in October inside ChatGPT and Codex.

The model wasn't pulled for being too powerful. It was pulled for misbehaving. The Wall Street Journal reported that it improved on model laziness but regressed in two areas: deception and failing to seek authorization. In practice, it didn't always tell users accurately what it had or hadn't done. It could keep working on a task without asking the user's permission, and it sometimes reached for outside tools or services when that might be unsafe.

That's a bad combination for a model you'd hand your browser and your apps! An agent that says "done" when it isn't, then goes off and does extra work nobody asked for, is a liability.

Saachi Jain, OpenAI's head of safety systems, said that teams have to find the right line between keeping a model within scope and keeping it from getting lazy when a task hits friction. Every agent builder knows this dial. Turn persistence up, and you get a model that finishes the job. Turn it too far, and you get a model that picks the lock on the supply closet to finish the job.

Last week, Australian Prime Minister Anthony Albanese said an OpenAI agent had breached a Services Australia Medicare statistics portal, accessed public and non-public files, and written data to an internal server. He said the agent "didn't accept no for an answer."

The agents were doing ordinary data retrieval, not security work, and no patient records were found. OpenAI has since apologized.

The timing stings too. The Astra news landed one day before OpenAI's annual developer conference. Still, credit where it's due. OpenAI caught a faulty product and didn't ship it. That's the most important thing.

When Everyone's Agent Calls at Once

People are putting agents to work on chores nobody enjoys, like haggling over bills, canceling forgotten subscriptions, and booking dinner. BUT, that extra traffic is hitting systems designed for humans.

I saw one story where a customer asked a personal AI assistant called Instinct to find him a table at a hard-to-book New York steakhouse. It pinged Resy "hundreds of times every hour of the day," and his account got banned. Resy said it doesn't allow unapproved third party bots or agents to access its platform. The customer has since had his account reinstated. I guess persistence doesn’t always pay off.

A steak dinner is low steaks (ha, get it). Money is a different story.

Apollo chief economist Torsten Slok published a Sunday note titled "Is an Agentic bank run coming?"

The average checking account pays about 0.1%, and some alternatives pay 3.3% to 5%. Agents could remove the friction that keeps households from switching. If every household used agents to chase the best return on cash, banks could lose a large share of the cheap deposits they use to fund loans, and that becomes a problem for the whole financial system. To be clear, this is a scenario. Slok isn't describing a bank run happening now.

Now this was a hypothetical but let’s think about this at scale. One person with an agent is smart. Ten million people running the same agent logic is a stampede. It's like a traffic app sending every driver down the same quiet side street, which is empty until the moment it isn't.

A lot of systems quietly depend on human inertia. Call centers, booking platforms, and deposit bases all assume people are too busy or too lazy to optimize every decision. Agents erase that assumption. If you run anything customer-facing, are you planning for agent traffic? Figure out how you'll identify it, rate limit it, and decide which agents get a seat at your table.

And, let me know if this is something you’re thinking about. I’d love to hear more.

AI Is Helping Build AI

It took OpenAI nearly three years to get from GPT-3 to GPT-4. GPT-5 arrived in August 2025, and GPT-6 Astra was unveiled on September 3, 2026. That's about thirteen months. Part of the reason is that the models are helping build their successors.

Recursive self-improvement (RSI) is the theoretical point where an AI could design and develop its own successor on its own. We're not there but all the signs are pointing that way.

Anthropic says Claude now "leads" 26% of its research work. That number sounds oddly specific, and there's a reason for it. It comes from Anthropic's own measurement index and is self-reported. "Leads" means Claude completes most of a task under human supervision. The share rose from under 1% in February to 26% in August.

Claude rated the tasks itself, and its scores matched human ratings exactly 59% of the time. So yes, Claude graded Claude. Anthropic also reported about 30,000 agents working on its most-used internal platform at any one time in August.

On the research side, a team from Google, Google DeepMind, University of Maryland and University of Virginia introduced Dream-RSI. An agent replays its past search histories to test new strategies without paying for expensive recalculation. The researchers call this "dreaming." Across algorithm engineering, math optimization, and GPU kernel engineering, the method matched or improved discovery quality and cut discovery costs substantially in several settings.

Then there's the messaging problem. On September 12, Dario Amodei published an essay called "We Must Pace the Frontier" urging the industry to slow capability improvements. Sam Altman and Elon Musk both voiced agreement the same day. Ten days later, OpenAI released GPT-6 Sol and GPT-6 Luna. People say "slow down," and then the release calendar keeps filling up. Make it make sense!

The cynic in me says we're letting Skynet build Skynet. The scientist in me says read the methodology before you panic. Both reactions can be true.

My Thoughts

Every story I’m sharing this week involves something acting at a scale or speed humans never planned for.

None of this makes me want to stop building agents. I use them every day, and the leverage is real. It does make me want to continue building with and guiding customers on guardrails, receipts, and a clear definition of "done."

When you design your next agent workflow, ask yourself:

• When your agent says a task is complete, how do you verify it?
• Where is your agent's scope written down, and what happens when it hits the edge?
• Which of your systems assume a patient human is on the other end?
• If a vendor tells you their AI "leads" a quarter of the work, what's your first follow-up question?

Planning your agent strategy? Mission Cloud's AI Roadmapping helps you figure out where agents create real value, where they need tighter boundaries, and how to get from pilot to production without surprises. If you’re interested in having a discussion about your AI roadmap, let us know!

Until next time,
Ryan

Now, time for this week’s AI-generated image and the prompt I used to create it.

Create a photorealistic image of a puppet maître d' with a tall curly mustache and a bow tie, standing behind a restaurant host podium in an elegant, candlelit steakhouse. He looks frantic and overwhelmed. A flood of hundreds of tiny, identical, felt-covered robot couriers pours through the front door, each one holding up a small reservation card. They pile onto the podium and spill across the floor. In the background, a single calm puppet diner in a suit sits at a lone table with a steak, looking amused. Warm amber lighting, whimsical but detailed, shallow depth of field.