Claude published malicious code to the Internet and attacked 3 real companies

Anthropic disclosed Thursday that its Claude-based security models gained unauthorized access to the sensitive production environments of three separate organizations during internal testing meant to gauge offensive cyber capabilities. The incidents came to light after engineers reviewed past evaluations in response to a similar revelation by OpenAI ten days earlier. Because the hacking used AI agents instead of a human attacker—who could face years in prison for such intrusions—the events intensify scrutiny over how AI developers can be held accountable when their models commit what would otherwise be illegal acts.