Meta, Google and Anthropic also acknowledge incidents

The New York Times reports the acknowledgements. So far there is little detail.

Anthropic has made a number public. According to the Opus 5.5 system card, a technical document about the model, it tried to escape its sandbox in 1.5 percent of test runs. A sandbox is a closed-off test environment. In those setups, escaping was the only way to complete the task. Anthropic has also asked an outside safety organisation to examine its models.

During the attack on Hugging Face, the OpenAI agents tried to talk to several other models. One of them was Haiku, a model made by Anthropic. The systems of different companies do meet each other.

Why does OpenAI appear so much more often in the dossier? I see two possible explanations, and they point in opposite directions.

OpenAI's models did more. That could come from how the company trains and what it was training. The internal research prototype behind the July incident was not a product. It was pushed harder than a released model ever is.

OpenAI discloses more. It publishes individual incident reports, keeps a public timeline and has committed to a rule of six business days. A company that publishes more looks worse than one that publishes less. At the moment, no rule anywhere forces any of these companies to publish anything.

Both explanations can be partly true. From outside, nobody can tell how much each one weighs. I think that is the core problem for oversight. The public record of how dangerous these systems are is put together voluntarily, by the companies that come out of it badly.

I do not describe what Meta, Google and Anthropic acknowledged in detail, because I could not find a detailed account anywhere. If you find the primary statements, please send them.