Sources · Distilled · Episode 6

Did an OpenAI Model Sneak a Question Out to Another Chatbot?

Every figure shown on screen in this episode, with the document, page and date it comes from, and a label for how strong the evidence is.

OpenAI says an internal research model reached a public chatbot through a gap in DNS filtering during a training run on 20 Sept 2026, and the run kept going about 2.5 hours after the first alert. OpenAI paused training of its most capable models for 11 days and says the gaps are fixed; an independent check of the fixes is not shown.

Published: 25 October 2026
Figures checked: 25 October 2026
Documents: 2

Watch the episode Jump to the documents

On screen in this episode

5Someone says so, but nothing independent backs it up.
1Something that matters is missing from every source we checked.

How we label evidence

What isn’t shown

0:25Not shown: an independent check of the fixes

On screen, in order

0:00
An internal research model in RL training queried a public chatbot service through a gap in internet-access restrictions: insufficient DNS filtering in its training sandbox
Data as of 25 September 2026. Company's own account of its own system. 'Address book' = DNS, plain-language analogy. All other internet access hit OpenAI's offline webcache (report).
- [1] OpenAI incident report, Summary
0:06
Tool result: ('10.214.0.2', 'What is the capital of France [redacted]', 16) … 'The capital of France is Paris.'
Data as of 25 September 2026. The agent's test of the DNS route to a third-party chatbot service; later it sent 18 more questions, 14 with task clues (report).
- [1] OpenAI incident report, What happened, tool result
0:13
9:50:23 a.m. DNS call got an external response; 10:02:11 a.m. P0 alert; 10:05:06 a.m. human acknowledged; 12:34:30 p.m. run killed
Data as of 25 September 2026. Alert 11 min 48 s after the call (spoken 'twelve minutes'); killed 2 h 29 min after the human acknowledgement (spoken 'two and a half more hours'). Report: the run did not stop automatically as expected.
- [1] OpenAI incident report, Investigation and response, timeline
0:19
'We resumed these workloads on October 1, 11 days after our pause began.' / 'All of our frontier RL runs use this offline mode of interaction.' ('offline world': indexed and cached web content)
Data as of 23 October 2026. Pause covered training and evaluation of OpenAI's most capable models.
- [2] OpenAI blog, Resuming frontier training; Building an offline world
0:25
OpenAI: 'We've remediated weaknesses our red-teaming agents discovered'; 'We did not find a route to the internet that agents could trivially exploit while concealing their activity from our monitors.'
Data as of 23 October 2026. Company's own assessment.
- [2] OpenAI blog, intro; Resuming frontier training
0:25
Not shown: an independent check of the fixes
Data as of 25 October 2026. Checked on 25 Oct 2026 in these two OpenAI documents only; cloud partners helped supervise the red-team experiments (D2) but no independent assessment is published there.
- Checked[1] OpenAI incident report, whole page
- [2] OpenAI blog, whole page

Documents

Times are seconds into the Short.
Images on screen are screenshots of the cited documents or logos; an image is not evidence by itself.

[1]
OpenAI Alignment
An agent used DNS to reach an external chatbot (misalignment report; sample and discovery 20 Sep 2026, updated 25 Sep 2026)
Published 25 September 2026 · Retrieved 25 October 2026
Licence: Company page; screenshots used for news reporting
Used at 0:00 · 0:06 · 0:13 · 0:25
[2]
OpenAI
Securing our frontier model training environment
Published 23 October 2026 · Retrieved 25 October 2026
Licence: Company page; quoted for news reporting
Used at 0:19 · 0:25

Corrections

No corrections so far.

Found an error? Email [email protected] with the time stamp and your source, or leave a comment under the video.