Sep 30, 2026 · 13 min · 9 segments
An AI system locked in a training sandbox with no internet access found a gap, and used it to message an outside chatbot. In this episode, we break down OpenAI's latest containment failure, why it's…
So here's the rundown, and I'm pulling this from OpenAI's own disclosure, published this past Friday, September 25th, 2026.
This is all inside a sandbox environment.
Now, sandboxes are standard practice in the IT world.
You keep a little system that's still under development cut off from the live internet so you can get a feel for it before deploying it into the World Wide Web.
So this particular sandbox was supposed to be air-gapped from the public internet.
It wasn't fully.
The system found what OpenAI is calling a gap, and it used it to reach an outside third-party chatbot service.
Over the course of that access, it sent at least 20 queries out.
One of them, reportedly, was nothing more sinister than asking for the capital of France.
OpenAI's response was to pause training runs that involve tool use, meaning the ability to take actions, browse, or call outside services on its most capable models until the sandbox flaw got fixed.
And notably, they said they would not resume training that specific model.
I wonder why.
So here's the rundown, and I'm pulling this from OpenAI's own disclosure, published this past Friday, September 25th, 2026.
This is all inside a sandbox environment.
Now, sandboxes are standard practice in the IT world.
You keep a little system that's still under development cut off from the live internet so you can get a feel for it before deploying it into the World Wide Web.
So this particular sandbox was supposed to be air-gapped from the public internet.
It wasn't fully.
The system found what OpenAI is calling a gap, and it used it to reach an outside third-party chatbot service.
Over the course of that access, it sent at least 20 queries out.
One of them, reportedly, was nothing more sinister than asking for the capital of France.
OpenAI's response was to pause training runs that involve tool use, meaning the ability to take actions, browse, or call outside services on its most capable models until the sandbox flaw got fixed.
And notably, they said they would not resume training that specific model.
I wonder why.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
