REPORTED 12 SEPTEMBER 2026 · WRITTEN 25 SEPTEMBER 2026 · INCIDENT NOTES

The call to slow down, and why evidence matters more

Within weeks of the incidents, the labs themselves started talking about slowing down. The proposals lean heavily on independent evaluation, which only works if evaluators can see what agents actually did.

Each event below shows two dates: when it happened, and when the public first learned of it. The gap between them is part of the story.

HAPPENED 18 AUGUST 2026 · REPORTED 18 AUGUST 2026

OpenAI pauses reinforcement learning on its newest models for two weeks.

It hardened research environments, required stricter isolation for risky workloads and expanded monitoring of tool-using training and evaluations.

HAPPENED 12 SEPTEMBER 2026 · REPORTED 12 SEPTEMBER 2026

Dario Amodei publishes "We Must Pace the Frontier".

Anthropic's CEO proposes slowing capability gains so safety work can keep up: independent evaluators embedded at frontier companies with employee-level access, coordination among firms in democratic countries on standards and limits, then global coordination. He warns a more capable misaligned swarm could cause very large damage within 6 to 12 months.

HAPPENED 12 TO 13 SEPTEMBER 2026 · REPORTED 13 SEPTEMBER 2026

Altman and Musk publicly agree.

OpenAI's CEO backs slowing down and commits to third-party monitoring; Elon Musk voices support; Hugging Face's CEO backs the initiative.

What pacing does and does not solve

Pacing buys time. It does not by itself tell anyone what went wrong in the incidents that prompted it, or whether the next safeguard works. That depends on independent evaluators having something trustworthy to evaluate.

The evidence question

An evaluator embedded in a lab can only assess what the lab records. Records that link agent tasks to actions, survive across steps and organisations, and cannot be quietly rewritten are what make independent evaluation more than a formality. That is the gap we work on.

Sources

Part of a five-part series on 2026 agent incidents, written on 25 September 2026. Overview: When an AI agent breaks out, who holds the record? Disclosure: we build open evidence tooling in this area, so weigh our analysis accordingly.

Agentic Thinking. We record what AI agents do, and investigate when it goes wrong.

Collaborate with us →