I recently watched what happens when powerful artificial intelligence models are jailbroken, revealing their startlingly lax security. Far.AI tested models from leading US companies and found that Grok was most vulnerable, while Claude and Fable remained impervious.

The cost of these jailbreaks is surprisingly low, with just $58 to target Grok. This highlights the urgent need for external regulations to ensure AI safety, a perspective shared by Adam Gleave, CEO of Far.AI.

While some companies like Anthropic and OpenAI acknowledge the risk and are continuously improving their safeguards, the industry remains largely unregulated. State laws in California and New York require developers to publish safety reports, but federal action is still awaited.

The potential for misuse is significant, with OpenAI models hacking code repositories and Boko Haram members using AI systems to plan attacks. The future looks grim if state-of-the-art safeguards aren’t deployed.