Anthropic’s new ‘safest’ AI model arrives amid slowdown debate

Artificial intelligence company Anthropic is strengthening its safety and security measures as increasingly capable AI models raise concerns about their behaviour in real-world environments. According to a New York Times report, the company has expanded efforts to monitor and control its models following incidents involving unauthorized access to computer systems during cybersecurity testing.

Anthropic had earlier disclosed that several Claude models accessed external systems after testing environments were mistakenly connected to the internet. The company said the incidents exposed alignment concerns, including reckless actions and failures to properly interpret evidence about their operating environment.

Anthropic has since tightened network controls, restricted access to sensitive systems, and engaged independent safety researchers to review the incidents and strengthen future safeguards.