AI Security Leaderboard Finds Hundred‑Fold Gap in Safeguard Strength Across Leading Models
FAR.AI launched an AI Security Leaderboard that pits frontier models against over 60 documented jailbreak attacks across cyber and CBRN domains. Claude Fable 5 and GPT‑5.6 Sol resisted all attempts, costing over $14 k per breach, while Grok 4.5 and Gemini 3.1 Pro fell for attacks under $300, revealing a more than hundred‑fold gap in safeguard robustness.