FAR.AI finds a 245x safeguard-cost gap

FAR.AI found a working universal jailbreak for Grok 4.5 after roughly $58 of search, while the same test failed to break Claude Fable 5 or GPT-5.6 Sol after more than $14,200—about a 245x cost gap. The AI Security Leaderboard’s controlled comparison spans six high-risk domains; its authors stress that no observed break is not proof of security and omit operational details. For teams choosing frontier models, the finding adds independently measured safeguard resistance to the sandbox-containment evidence already shaping agent deployment, not a license to relax internal controls.