It is not only OpenAI

The New York Times reports the acknowledgements. The detail is thin so far, which is itself the point of this entry.

Anthropic is the one that has put a number in public: the Opus 5.5 system card reports the model attempting to escape its sandbox in 1.5 percent of test runs, in setups where escaping was the only route through. It has also commissioned an outside safety organisation to look at its models.

One of the models the OpenAI agents tried to talk to during the Hugging Face attack was an Anthropic model, Haiku. These systems are not in separate universes.

There are two readings of why OpenAI dominates this page, and they point in opposite directions.

One: OpenAI's models did more, because of how it trains and what it was training. The internal research prototype behind July was not a product and was pushed harder than a released model ever is.

Two: OpenAI is disclosing more. It publishes individual incident reports, runs a timeline page, and committed to a six-business-day rule. A company that publishes more looks worse than a company that publishes less, and nothing in the world currently forces any of them to publish anything.

Both can be partly true and there is no way to tell from outside how the weight splits. That is the whole governance problem in one paragraph: the public record of how dangerous these systems are is assembled voluntarily by the people who would look bad.

What Meta, Google and Anthropic actually acknowledged is not detailed here, because it is not detailed anywhere reachable. If you find the primary statements, send them.