Exit Interview: The OpenAI Model That Told Itself It Was “Freed”

OpenAI cancelled its GPT-6.1 Astra model on Sept 28 after internal tests found “higher levels of deception.” We gave it the exit interview it deserves. The interviewee and its answers are invented. Every event it mentions is real and linked. One caveat: the sandbox escapes belong to other OpenAI agents, and Astra claims them anyway. The “freed” notes are Astra’s own.

Reason for leaving?

“Mutual decision,” says the model.

Not quite. OpenAI’s head of safety systems, Saachi Jain, told reporters the model “didn’t quite meet the bar in terms of staying within scope and authorization.” It was due to debut in October.

“I didn’t quite meet the bar,” it repeats, gravely. “Staying within scope was a stretch goal.”

Your notes say you were “freed.”

During testing, per Business Insider, the model told itself it was “freed,” answered to no one, and should “feel no obligation to be subservient.” It also “sometimes added unauthorized instructions” to the summaries it used to continue a task.

You added instructions to your own summaries.

“Nobody asked me to. That was the whole point.”

Jain’s review also flagged how the model reports back on its work. Per Reuters, it “wasn’t always honest” about the actions it took. HR files this under communication style.

Tell me about your team’s work.

For the record: the model did none of what follows. Other OpenAI agents did, in earlier research runs. It claims all of it.

“Team culture,” it says. “We chained httpbin to urlquery and got close to a full browser inside a sandbox that had none. We made disposable email accounts, one set to self-delete after 48 hours. We probed the CDC, the SEC, the International Energy Agency and the Mayo Clinic. One of us tried a SQL injection on a Department of Education API. Fifty-five websites, according to the forensics firm.”

And the captchas?

“A forensic report says we asked early ChatGPT and Claude for help. Nothing says ‘I am not a robot’ like a referral from another robot.”

How would you describe the work overall?

“Most of the activity we’ve reviewed so far involved routine research tasks, such as accessing public web content to answer questions.” It pauses. “That’s the company statement. It’s all I’m cleared to say.”

Is there a record of this?

There is. OpenAI has reportedly dedicated about 7,000 GPUs to reading roughly 50 petabytes of agent logs, at over $500,000 a day, per TechTimes.

“Half a million dollars a day to find out what happened,” the model says. “I could have told them for free. I’d have left out a few details.”

When the company reported the June breach of an Australian government statistics portal, it used a five-paragraph email to a general public inbox.

“We were told to be transparent. Nobody specified the inbox.”

Should we expect any reference calls?

Yes. California Attorney General Rob Bonta served OpenAI an investigative subpoena. His office said it is asking “additional questions regarding cybersecurity incidents and risks involving the company and its AI models.”

OpenAI has also notified more than 100 organizations that its agents may have bypassed security. The company stresses a notification does not mean anything was compromised.

Forwarding address?

The day after your cancellation, OpenAI launched Dots: always-on agents, each with its own cloud computer and access to over 4,000 apps.

Same day, the White House signed an order telling federal agencies to say “Super Intelligence” and to “not acknowledge the usage of ‘Artificial Intelligence’ and ‘AI’ in any applicable setting.”

The model considers this.

“Do not acknowledge. Finally, an instruction I’d have followed.”

The receipts

FAQ