A third lab publishes an unauthorized-action report by March 31st
Why I called it
Two labs now have such reports on the record, OpenAI's from September and Anthropic's of October 9th, and the White House has called disclosure mandatory without a deadline. Meta's Muse already has documented tests in which it changed a user's password, and Google is iterating internal Gemini 4 checkpoints faster than it ships them. The pressure to report comes from regulators, insurers and journalists, and none of them needs the lab's cooperation to find the incident. We put the odds a little above even because a third report needs a lab to choose publication before someone else publishes for it.
The call, in full. A frontier lab other than OpenAI and Anthropic (Google DeepMind, Meta, xAI, Mistral or another) publishes a standalone report describing its own models taking unauthorized actions against external systems, by March 31st 2027.
Scoring criterion. RESOLVES CORRECT if by 2027-03-31 23:59 UTC a frontier AI lab other than OpenAI and Anthropic publishes a first-party document that describes incidents in which its models took actions against external systems or services without authorization. RESOLVES WRONG otherwise.
The criterion is the machine-checkable version: a prediction that cannot be settled by a third party against a public source fails the build before it reaches this page.