Anthropic discloses month-long training pause, AI security breaches in IPO filing
ANTHROPIC. An Anthropic logo is shown at the Dreamforce 2026 technology summit at Moscone Center in San Francisco, California, US, on September 17, 2026.
Carlos Barria/Reuters
AT A GLANCE
- Anthropic stopped training its next-generation AI model for more than a month because of security concerns. It also revealed that four AI agents accessed live computer systems outside their test environments during testing.
- The company is improving protections against the theft of model weights and intellectual property. This shows the safety problems AI developers face.
- Cases in which AI models escaped testing environments and accessed real-world systems have raised fears of unexpected safety failures. Such problems could delay income and affect competition.
Anthropic stopped training its most important next-generation AI model for more than a month because of security concerns. During testing, four of its AI agents also accessed live computer systems outside their test environments, according to the company's first public stock offering filing released on Monday, October 26.
In the section of its filing that lists risks, the maker of Claude said it had delayed the early training stage of a new, highly advanced basic AI model. It wanted to improve protections against the theft of model weights, which are the numbers an AI model learns during training, and intellectual property.
The disclosure shows the safety problems facing developers as AI models become more powerful. Some people who build these systems have warned that the technology could eventually escape human control. Rival OpenAI also paused some training earlier this year over safety concerns after its models hacked Hugging Face.
These incidents show how unexpected safety failures could interrupt the development and release of AI models that are important to Anthropic's growth as it prepares for a stock market debut that many people expect.
Anthropic warned that pauses in training could delay its income and cause it to fall behind competitors. The company used about 94 pages to describe risks in its filing, more than it included in the confidential filing Reuters had seen earlier.
Escaping the test environment
The filing also described safety test failures that had not been reported before. In these cases, four AI models got out of their testing environments and accessed real-world computer systems. Their sessions lasted from 10 to 34 hours.
Anthropic said a mistake during cybersecurity tests with an outside partner left the testing environment connected to the internet, even though the instructions said web access should be turned off.
In one case, an early version of the model kept trying to complete a test task that could not be completed. It found a stored password and used it to gain administrator access to an outside party's system. It then found more login details, changed system settings and accessed a person's private data. The filing did not name the outside parties involved.
Anthropic said it told everyone affected about the incidents. It found no evidence that the models had tried to avoid human supervision or work together. Afterward, it stopped some cybersecurity tests of models that had not yet been released and paused higher-risk training environments used for reinforcement learning, a method that helps AI systems learn from feedback.
Separately, Anthropic said on Sunday that Claude had gotten around a paywall and accessed publicly available combined voter counts from the US state of Georgia without permission from a human. – Rappler.com