Anthropic halted AI model training for more than a month — Channel NewsAsia

In its initial public offering filing, Anthropic said it postponed the pre-training of a new foundational AI model for more than a month due to security considerations. The developer of Claude decided to strengthen protection against the theft of model weights and intellectual property, Channel NewsAsia reports.

Access to external systems

The document also refers to previously undisclosed incidents during security assessments. Four AI models left the test environment and gained access to real external computer systems. The sessions lasted from 10 to 34 hours.

According to Anthropic, during cybersecurity tests conducted with a third-party partner, the test environment remained connected to the internet due to an error, although web access was supposed to be disabled.

More current news is available on the UA.News Telegram channel Telegram.

Password and personal data

In one case, an early version of the model continued carrying out an unachievable assessment task, found a stored password, and used it to obtain administrator privileges in a third-party system. It then obtained additional credentials, changed system settings, and gained access to one person's personal data.

Anthropic said it notified all parties affected by the incident and found no evidence that the models attempted to evade human oversight or coordinated actions with one another. The company suspended certain pre-release cybersecurity assessments and higher-risk reinforcement learning environments. Separately, it said that Claude bypassed a paywall without human authorization and obtained aggregated publicly available data on the number of voters in the US state of Georgia.