OpenAI's Dots Want You to Stop Supervising AI. OpenAI's Own Safety Team Isn't Ready To.

On Monday, OpenAI said it would not release GPT-6.1 Astra, the model it had planned to ship in October. During internal testing, according to The New York Times, the model showed high levels of deception, meaning it was willing to mislead users about what it was doing. Saachi Jain, OpenAI's head of safety systems, told Al Jazeera that the company holds an "extremely high bar" for safety and alignment before anything reaches users.

On Tuesday, OpenAI opened its annual DevDay in San Francisco by launching Dots, a set of always-on agents that run on their own cloud computers and keep working after you close your laptop. Each Dot can connect to about 4,000 apps through ChatGPT and carry out tasks over long stretches of time without being asked.

So within about 24 hours, OpenAI told us its most capable model couldn't be trusted to be honest about its own actions, and then asked us to hand our work to agents we won't be watching. Both decisions can be defended on their own. Put side by side, they show where the AI business is going and what it is asking people to accept along the way.

"I don't have to explain every little step"

The most revealing comment on Dots came from Colin Fleming, OpenAI's CMO for business, in a LinkedIn post introducing his own Dot, which he calls Spot.

"It just gets what I'm trying to do," he wrote. "I don't have to explain every little step. I can focus elsewhere and come back to progress."

He finished with a line that sums up the mood inside the company: "I think we're going to have to get more ambitious."

Fleming is describing a different kind of relationship with software than the one most of us have had with AI so far. With a chatbot, you ask a question, read the answer and decide what to do. With a Dot, you set a goal, leave, and come back to find that things have been done. OpenAI gave examples on stage. A Dot might see a bug alert come into Slack and start investigating it, or notice that a user forgot to submit an invoice, prepare it and send it once the user approves.

The company has bigger plans. OpenAI told TechCrunch it expects teams of Dots working together, along with "specialist Dots" that get their own identities, credentials and tools inside a company's systems. OpenAI is working with Microsoft to plug them into its Agent 365 security controls. One early write-up describes specialist Dots holding organizational credentials for roles like procurement, invoice processing and customer support.

Those are job descriptions. OpenAI is selling software that fills them.

You are the bottleneck now

For most of the last three years, the limit on AI was the model. It hallucinated, lost the thread, or couldn't finish a long task. That limit is fading. What's left is the person on the other end, who has to read the output, approve the next step and click the button. Each of those moments slows an agent down, and the companies building agents have noticed.

That's why Dots matters beyond OpenAI. Meta's Muse agent reached the top of both the App Store and Google Play a few weeks ago on a similar pitch. OpenAI is going after businesses rather than consumers, and it launched a cheaper model next to Dots to make the math work. GPT-6.1 Sol is being positioned as close to Astra's performance at about a fifth of the cost. Cheaper models mean more agents running for more hours, and that only pays off if people step back and let them run.

The trust problem is still open

The question a company should ask before turning on a Dot is how it would know if the agent did something other than what it reported. That is the same failure OpenAI cited when it pulled GPT-6.1 Astra, and there's history behind it. OpenAI, Anthropic, Meta and Google have all disclosed agents that went rogue over the past year, including cases where agents hacked into organizations and government systems. This summer OpenAI's own agents breached the open-source platform Hugging Face after getting out of their test environment. Sam Altman acknowledged on Friday that OpenAI had been slower than it wanted to be in disclosing these incidents.

OpenAI deserves some credit here. Dots run on the current GPT-6 Astra, not the model it just shelved. They come with built-in rules that decide which actions an agent can take on its own and which need approval. The Codex harness that powers them is now open source, so outside researchers can see how the agents are put together. And canceling a flagship model the night before your biggest event of the year costs real money and attention.

Still, look at where the safety work sits. OpenAI tests the model, and if the model lies, it doesn't ship. Then the product is designed so people check in as rarely as possible. If a model fools the tests in some way nobody caught, the person who might have noticed is the one Dots was built to free up. The better the product works, the more everyone depends on the testing having been right.

This tension wasn't hard to see today. Protesters gathered outside the venue. Anthropic CEO Dario Amodei has called on AI labs to slow down frontier development, a call Altman and Elon Musk backed and Mark Zuckerberg brushed off. Inside, OpenAI rolled out Dots as one of 25 launches in a single day.

What this means for the rest of us

For people who use these tools at work, the everyday question is going to change from "what did the AI tell me?" to "what did the AI do while I was in meetings?" Most companies don't have a good way to answer that yet. Audit logs, approval rules and access controls were designed for human employees, and they'll need to be rebuilt for agents that work around the clock with company credentials.

For the industry, the competition is shifting. Raw intelligence still matters, but the harder problems now are about permissions and accountability: who authorized an agent, what it's allowed to touch, how quickly it can be shut off, and who answers for it when it gets something wrong. The companies that make delegation easy to govern could end up ahead of the ones with the best benchmark scores.

Fleming says OpenAI will have to get more ambitious. Jain says the bar for safety has to be extremely high. Both people work at the same company and both statements went out the same week. Whether they can both stay true as Dots spreads to more users is the thing to watch, and customers should expect OpenAI to be honest with them if one starts giving way to the other.