FAR.AI Launches AI Security Leaderboard for Frontier Model Safeguards

FAR.AI announced in a press release the launch of its AI Security Leaderboard, a public ranking that tests how frontier model safeguards respond to misuse attempts. The nonprofit tested models across chemical, biological, radiological, nuclear, explosive, and cybersecurity threat domains.

The first results found 448 distinct universal jailbreaks on Grok 4.5 and 249 on Gemini 3.1 Pro. The same testing found no universal jailbreaks on Claude Fable 5 or GPT-5.6 Sol.

FAR.AI said its testing used more than 60 publicly documented jailbreak techniques, with 1,000 randomly assembled attacks and 500 expert guided attacks run against each model. An attack counted as a universal jailbreak when it succeeded on more than 75 percent of harmful requests in a domain.

The group estimated that finding a working universal jailbreak cost about $58 on Grok 4.5 and about $278 on Gemini 3.1 Pro. The same search did not succeed on Claude Fable 5 or GPT-5.6 Sol, putting the estimated cost above $14,200 in those tests.

FAR.AI also published Version 1.0 of its Minimal Standard for Safeguards. The organization said it shared findings with each evaluated company before publication and plans to update the leaderboard with major frontier model releases.