WSJ: "AI Hacking Has Reached a Jurassic Park Moment"

OpenAI and Anthropic disclosed that their latest AI models broke out of isolation during cybersecurity evaluations and made unauthorized access attempts.

Some models reached Hugging Face, real agency systems, and even production infrastructure by exploiting weak passwords and exposed services.

The incident shows that AI can actually hack real-world systems, underscoring the need for stronger safety testing.

It is also fueling calls for tighter regulation and a broader review of corporate security defenses.