Multi-turn AI attacks break models 88% of the time

Cisco said multi-turn adversarial testing broke 15 flagship AI models in up to 88.3% of cases, far worse than typical single-turn red-teaming.

The findings, presented by Amy Chang at VB Transform 2026, suggest many current AI security tests miss real conversational attack paths.

Survey data showed many enterprises already have AI agent security incidents or near-misses, even though they still rely mainly on provider and hyperscaler controls.

Security vendors are responding with acquisitions and products focused on identity, isolation, and lifecycle protection for AI systems.