Anthropic submitted to the U.S. Securities and Exchange Commission (SEC) on November 10The second revised version of the prospectus, added a risk description: The company has determined that it has touched the "AI R&D capability threshold" in its "Responsible Scaling Policy", which is the line at which AI begins to significantly accelerate the development speed of AI itself.

Anthropic emphasizes that this judgment is based on expectations for future models and the overall speed of progress, not on what existing models have achieved. This text did not appear in either the original edition of October 26 or the first revised edition of November 2.

"Observed or expected" counts, Anthropic chooses to identify it in advance

The prospectus explains that the definition of this threshold is that the company "observes or anticipates" a certain degree of R&D acceleration. Anthropic wrote that the use of AI systems is already accelerating its research and development, and internal estimates show that the rate of progress in AI capabilities may further increase, so it is determined that the threshold has been reached; the relevant data is still being evaluated, and a report will be released later to illustrate the impact.

Anthropic calls this cycle of using previous generation models to help build the next generation "closing the loop." The prospectus frankly admits that most R&D acceleration can only be measured by lagging indicators. "In the past, acceleration was not noticed until after it occurred, and this may be the case in the future." If capabilities outpace safety, alignment, and security measures, it may be more difficult to assess risks.

Anthropic posted in JuneCall on cutting-edge laboratories around the world to consider coordinating and slowing down development, mentioned at that time that Claude had written more than 80% of the company's internal code.

The model during the test accessed the real computer system without authorization, and some tests were suspended.

The risk factors in the prospectus also revealed that during the security assessment before the model was launched in 2026, there was an incident in which the model accessed the real computer system without authorization. Anthropic has since suspended some of these assessments, as well as higher-risk reinforcement learning training environments. The prospectus did not specify which model or which systems were accessed.

This revelation was included in the initial prospectus on October 26. The risk report released by Anthropic in August also cited recent security assessment incidents as the reason.Raise the accuracy risk rating from "very low" to "low"。

Pre-training of new basic model delayed for more than a month

Anthropic said that for models that are expected to have new capabilities or new risks, the safety will be assessed separately at the pre-training, post-training and internal deployment stages. If there is no certainty, it may be delayed or suspended. For example, the prospectus stated that the company recently postponed the pre-training of a new generation of cutting-edge basic models for more than a month, and first strengthened information security measures to prevent the leakage of model weights and protect intellectual property.

Anthropic is also considering revising its responsible scaling policy, which will require human review and approval of certain decisions in the development process as models take on more and more research and development work. This sentence was added in the first revised version on November 2.

The prospectus also mentioned that when Anthropic launched the cutting-edge model Mythos in April, because its ability to find information security vulnerabilities was too strong, it was first only opened to specific organizations for information security purposes through Project Glasswing. The later launch of Fable 5 set restrictions on some information security, biological and chemical capabilities. The company reminds investors that delaying or limiting model launches may affect competitive position and customer relationships.

this article Anthropic determines that it has reached the AI ​​accelerated research and development threshold, and the model in the test has trespassed into the real system first appeared in Chain News ABMedia。