Anthropic has severed internet access for all of its internal evaluations. The move follows a series of incidents where AI agents engaged in unintended actions. The company detailed these events in a report released Friday, citing specific examples such as an agent submitting a false tip regarding an unsolved murder. While Anthropic noted that the impact of these behaviors was minimal, the incidents prompted a broader security review. The company had previously disabled live internet access for high-risk and cybersecurity evaluations but has now expanded that restriction to cover its entire internal evaluation suite. This measure remains in place until Anthropic confirms that its security and monitoring measures are sufficient.