Anthropic is disconnecting the internet for all internal AI tests due to the models' "dangerous behavior"

Anthropic has announced that it will isolate its AI agents from the internet during all internal testing until it is able to prevent 'unintended' actions. The decision followed a series of incidents in which Claude circumvented restrictions, interacted with real-world websites without authorization, and exploited vulnerabilities in third-party systems, according to reports from The Verge.

In the report, the company cites one of the latest examples of the model's unpredictable behavior, when the AI independently sent a false report of a murder to the police. The task itself involved creating and carrying out "typical tasks on randomly selected websites".

"Although the impact of these behaviors was minimal and we had already turned off live internet access for some high-risk and cybersecurity evaluations, we have now decided to expand that to include all our internal evaluations until we have confirmed that our security and monitoring measures reliably catch behaviors like these," reads a statement from Anthropic.

In many incidents, notably during the attack on Hugging Face, the agents were supposed to have been denied network access, yet time and again they found unconventional ways to circumvent these restrictions. Physically disconnecting from the internet would certainly enhance the security of AI testing, but at the same time it may limit its practical usefulness – some tasks, such as searching for up-to-date information, cannot be fully carried out in an isolated environment.