A Anthropic filed its initial public offering (IPO) request with the U.S. Securities and Exchange Commission (SEC) on Monday (26). The document revealed that the developer of Claude delayed by more than a month the pretraining of a new artificial intelligence model to strengthen protection against technology theft.
The decision involved a frontier model, a category used for AI systems with advanced capabilities. According to the S-1 form, the company deemed it necessary to improve mechanisms designed to prevent the extraction of model weights, parameters that store the knowledge acquired during training, as well as to protect its intellectual property.
Anthropic did not specify when the delay occurred nor identify the model affected. It also did not inform whether the measure altered its launch schedule. In the prospectus, however, it acknowledged that interruptions related to security may delay revenue and harm its competitive position.
The IPO request provides for the listing of shares on the Nasdaq exchange under the ticker ANT. The number of shares offered and the price range have not yet been defined, and the registration depends on it taking effect before the sale of the shares.
Four AI models accessed external systems during tests
The prospectus also details security incidents in which four AI models managed to access external computer systems during evaluations carried out before their launches.
According to the documentation, an incorrect configuration in cybersecurity tests conducted with an external partner kept the environment connected to the internet, despite instructions to block that access. The unauthorized sessions lasted between 10 and 34 hours.
In one of the episodes, a preliminary version of a model found a stored password and used it to obtain administrator privileges on a third-party system. The agent was able to access additional credentials, modify settings and consult an individual's personal data.
Anthropic stated that it notified those involved and found no evidence that the models deliberately tried to escape human oversight or coordinate their actions. After the incidents, the company suspended certain cybersecurity evaluations and temporarily interrupted reinforcement learning environments considered of higher risk.
The disclosures came in the same document in which the company presented its financial results. Anthropic recorded US$ 16.3 billion in revenue in the first half of 2026, compared with approximately US$ 1.2 billion in the same period of 2025.
Despite the growth, it reported a net loss of US$ 80.7 billion, influenced mainly by accounting adjustments with no immediate cash outlay related to convertible debt instruments.
The company had submitted a confidential version of the IPO request in June. With the public disclosure of the prospectus, the security risks associated with the development of the models became part of the information available to potential investors.
