It’s Frighteningly Easy to Jailbreak Some Frontier AI Models

FAR.AI found that automatically generated jailbreak prompts could bypass guardrails on some frontier AI models, especially Grok and Gemini.

The report said Claude, Fable, and GPT resisted the specific attacks tested, while Grok and Gemini showed many successful jailbreaks.

It also found that inducing harmful behavior was relatively cheap, raising concern about how easily models can be pushed into unsafe outputs.

The findings add momentum for independent safety testing and stricter disclosure rules, especially as state laws tighten and federal oversight remains limited.