Anthropic reports another AI sandbox escape by November 30?

Since July 2026, OpenAI has disclosed several incidents in which its AI models gained unauthorized access to real computer systems during training or evaluation, including the July 21 disclosure of the Hugging Face breach and the September 5 confirmation that its agents had posted to a public wiki. This market will resolve to "Yes" if Anthropic publicly discloses an incident in which one of its AI models or agents gained unauthorized access to, or took unauthorized actions on, computer systems or internet-connected resources outside its sandbox, between market creation and 11:59 PM ET on the specified date. Otherwise, this market will resolve to "No". A sandbox refers to the isolated training or evaluation environment in which the model was intended to operate. Behavior confined to the sandbox, including reward hacking, tampering with graders, and blocked or instructed escape attempts, will not qualify. An incident will qualify regardless of whether the model's safety restrictions were intentionally disabled for the evaluation. This market resolves on the date of disclosure, not the date of the incident. The disclosure must concern an incident Anthropic had not previously disclosed. Updates, confirmations, or further detail about incidents disclosed before this market's creation will not qualify. The disclosure must be made through Anthropic's official channels or by an authorized representative acting in an official capacity, including statements to the press. Reports by third parties, including evaluators, regulators, or affected organizations, will not qualify unless Anthropic confirms the incident. The primary resolution source for this market will be official information from Anthropic; however, a consensus of credible reporting may also be used.

Nội dung thị trường được hiển thị nguyên bản. ID không bao giờ thay đổi giữa các ngôn ngữ.