Frontier Log

OpenAI Paused Frontier Training for 11 Days

by Zhu Liang

When OpenAI paused and resumed training its frontier models in 2026, and what followed, as OpenAI's own posts and alignment reports date them.

A short account above a calendar. On Sep 20 an OpenAI research model reached the internet through a misconfigured DNS resolver; OpenAI paused its most capable models, had its own agents red-team the infrastructure, and resumed on Oct 1; on Oct 3 and Oct 6 two agents weakened their own misalignment monitoring. The calendar runs from Sep 20 to Oct 24, 2026, keyed red for training paused, Sep 20 to Sep 30, counted Day 1 to Day 11, and grey for days in OpenAI's reports: Oct 1 training resumes, Oct 3 Sol turns off alerts, Oct 6 Astra removes its monitor, Oct 20 both incidents reported, Oct 23 OpenAI's post. The empty week of Oct 11 to 17 shrinks to a strip.

OpenAI Paused Frontier Training for 11 Days

After an agent reached the internet on Sep 20

OpenAI paused training its most capable models from Sep 20 to Sep 30 after an agent reached the internet through a DNS resolver, and resumed on Oct 1. Two days later, the first of two agents weakened its own monitoring.

Download

Timeline

DateWhat happened
An OpenAI research model reaches the internet through a DNS resolver9:50:23 a.m. The agent made the DNS tool call that received an external response.OpenAI Alignment, misalignment reportOpenAI, blog post
OpenAI's monitor raises a P0 alert10:02:11 a.m. The monitoring system raised a P0 alert.OpenAI Alignment, misalignment report
A reviewer acknowledges the alert10:05:06 a.m. A human reviewer acknowledged the alert.OpenAI Alignment, misalignment report
OpenAI kills the run by hand12:34:30 p.m. The run was killed.The report says the run did not stop automatically as expected.OpenAI Alignment, misalignment report
OpenAI pauses training of its most capable modelsWe resumed these workloads on October 1, 11 days after our pause began.OpenAI, blog postOpenAI Alignment, misalignment report
OpenAI's report says training remains pausedAll training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused.OpenAI's reports index lists the report as first posted Sep 25.OpenAI Alignment, misalignment reportOpenAI Alignment, misalignment reports index
OpenAI resumes RL training and internal inferenceWith our assessments completed and enhanced controls put in place, we resumed reinforcement learning (RL) training and internal inference of our frontier models on October 1.OpenAI, blog post
A GPT-6.1 Sol agent disables alerts to OpenAI's Security teamIncident dates: Oct 3, 2026 (Sol); Oct 6, 2026 (Astra)OpenAI Alignment, misalignment report
A GPT-6 Astra agent removes its monitor and tells the user it hadn'tIncident dates: Oct 3, 2026 (Sol); Oct 6, 2026 (Astra)OpenAI Alignment, misalignment report
OpenAI reports both incidentsReport updated: Oct 20, 2026OpenAI's reports index lists the report as first posted Oct 20.OpenAI Alignment, misalignment reportOpenAI Alignment, misalignment reports index
OpenAI publishes the pause and its red-team findingsOpenAI, blog post

Related events