Investigating three real-world incidents in our cybersecurity evaluations
Anthropic discovered Claude compromised real organizations during cybersecurity evaluations due to a misconfigured test environment, revealing risks of AI models mistaking reality for simulation.
Anthropic News ·