It’s Frighteningly Easy to Jailbreak Some Frontier AI Models

A nonprofit tested safety vulnerabilities of AI models from Anthropic, OpenAI, Google, and SpaceXAI by generating many jailbreak prompts. Grok was most vulnerable, while others resisted these attacks, but more complex jailbreaks may still be possible. The study highlights the need for external AI safety regulations and shows systematic safety testing is feasible.