FAR.AI Leaderboard Exposes Hundredfold Security Gap in Frontier AI

The Berkeley-based organization tested four leading models against a battery of chemical, biological, and cyber-threat scenarios. While Claude Fable 5 and GPT-5.6 Sol proved resilient, Grok 4.5 and Gemini 3.1 Pro showed significant vulnerabilities. Researchers successfully identified hundreds of universal jailbreaks in the latter two models using a systematic testing toolkit. The cost to find a working exploit was approximately $58 for Grok and $278 for Gemini, whereas the effort required to breach the more robust models exceeded $14,200.

Adam Gleave, CEO of FAR.AI, noted that the disparity is wider than industry expectations. The organization argues that current failures are not due to exotic, unknown threats but to a lack of basic defense-in-depth engineering. To address this, FAR.AI has released a Minimal Standard for Safeguards, setting a baseline for what frontier models should withstand. The leaderboard is intended to provide policymakers and the public with transparency, as the current "shop around" reality allows attackers to bypass one company's refusal by simply redirecting their queries to a less secure competitor.