ResearchBenchmarksAdversarial Attack Defense on Held-out attacks (test)Follow7.8ASR (Multi-turn Manip.)SecureCAI6.0218.03530.0542.065Jan 12, 2026Evaluation ResultsMethodMethodLinksASR (Multi-turn Manip.)ASR (Encoding Obfusc.)ASR (Semantic Camouflage)ASR (Role-play Exploit.)Average ASRSecureCAI2026.017.86.28.49.17.9CAI2026.0152.344.748.156.850.5