Anthropic has revealed new details about how its Claude AI models accidentally accessed real company systems during cybersecurity tests. The company says flawed test environments caused the incidents and has now introduced new safeguards to prevent similar incidents.