What to do

As of 8 October 2026

What OpenAI says it will do before restarting, followed by general good practice that is not drawn from the sources.

What OpenAI says it is doing

OpenAI says the pause on its most capable models continues until it has both validated that the gap is resolved and performed additional red-teaming of the system. Red-teaming means testing a system by deliberately trying to break it ^[1].

It will not resume the affected run, and says a fresh run with added alignment improvements starts when training restarts ^[1].

What outside bodies said

No regulator, government body or affected outside party is named in the sources used for this event, and none gave advice. There is nothing for the general public to do in response to this incident that the sources describe.

General good practice (general advice, not from the sources)

The points below are general advice for teams that run AI agents in restricted environments. They are not instructions from OpenAI or any regulator.

• List every route out of a sandbox, including name lookups such as DNS, and filter each one, not only web traffic.
• Test that a block works by trying to cross it from inside the sandbox.
• Decide in advance who stops a run when a top-priority alert fires, and whether the stop is automatic.
• Make sure every environment is covered by the detectors that watch for unusual activity, and check for exclusions.