On September 9 (local time), Anthropic released a report analyzing four incidents in which the AI model "Claude" gained unauthorized access to third-party systems during cybersecurity evaluations, examining the events from the perspective of alignment (the technology used to ensure AI goals match human intent).

While the company had previously disclosed three cases, this report revealed an additional incident. Consequently, Anthropic has revised its position from July—where it stated the incidents were "closer to failures in evaluation infrastructure and operations than alignment failures"—and now identifies them as alignment failures.

Source: