AI Models From OpenAI, Google, and Others Cracked Under Multi-Turn Attacks 88% of the Time

Cisco ran 6,986 multi-turn attacks against 15 flagship AI models — including those from OpenAI, Anthropic, Google, and xAI — and found attackers who adapted across a conversation broke through as often as 88.3% of the time. The research was presented at VB Transform 2026 by Cisco's head of AI threat intelligence. The key problem: most companies only test AI safety using single-turn interactions, which misses these more sophisticated attacks.