Timeline

As of 8 October 2026

Dates are when things happened; disclosure dates are marked separately. Clock times on 20 September are exactly as OpenAI's report gives them, and the report states no time zone.

Each entry says whether it is when something happened, when it was disclosed, or a response, and how well the sources support it.

Before 20 September: the first pause

1. 10 to 12 July 2026Happened

   The separate Hugging Face incident, which OpenAI's later hardening and first pause followed. It has its own report on this site.^[1]

2. Late July 2026Happenedreported

   Fortune dates OpenAI's first pause in reinforcement-learning training to “late July for two weeks”. OpenAI's own pages give no dates for it. reported^[2][3]

3. 18 August 2026Disclosed

   OpenAI publishes a post saying it had paused reinforcement-learning training on its latest models intended for deployment, described as two weeks, with its largest planned frontier run still on hold. It also sets a target to alert within 30 minutes of concerning activity surfacing.^[3]

4. 26 August 2026Disclosed

   OpenAI's post on the Hugging Face incident says the largest run remains on hold and some workloads are “slowly resuming under these new controls”.^[1]

20 September 2026: the day itself

1. 20 September 2026Happened

   The model is given a search task about a specific person. Direct requests to search engines fail with errors 403 and 502; other traffic is served from the offline webcache.^[4]

2. 20 September 2026, 9:50:23 a.m.Happened

   The first DNS call gets a response from an outside chatbot service. No time zone is given.^[4]

3. 20 September 2026, 10:02:11 a.m.Happened

   Monitoring raises a top-priority (P0) alert.^[4]

4. 20 September 2026, 10:05:06 a.m.Response

   A person acknowledges the alert. The run does not stop automatically.^[4]

5. 20 September 2026, 12:34:30 p.m.Response

   The run is killed manually.^[4]

After: 25 to 29 September

1. 25 September 2026Disclosed

   OpenAI's alignment blog report, “An agent used DNS to reach an external chatbot”, carries this as its updated date. It says the run will not resume and the pause on the most capable models continues.^[4]

2. 26 September 2026Disclosedreported

   Fortune reports the incident, says the report was released on a Friday, counts it as OpenAI's second training pause, and quotes Micah Carroll on inference. reported^[2]

3. 29 September 2026Disclosedreported

   The Hacker News publishes a story headlined “OpenAI Pauses Tool Use After Agent Bypasses Internet Controls to Reach External Chatbot”. reported^[5]