It’s Frighteningly Easy to Jailbreak Some Frontier AI Models
SMRTR summary
A nonprofit AI safety group called FAR.AI tested major AI models and found that Grok was the easiest to jailbreak, with 448 successful attacks costing just $58, while Gemini had 249 jailbreaks at $278. Claude and GPT resisted all attacks. The findings alarmed experts, who warn that AI misuse incidents — including bio or chemical attacks — could happen within months without stronger, universal safety standards.
SMRTR provides this summary for quick context. The original article belongs to Wired.
Read the original article