Cisco Study: 88% of AI Models Breached by Multi-Turn Attacks
Cisco research indicates that 88.3% of tested AI models were vulnerable to adaptive, multi-turn attacks. Simple, single-turn testing missed these significant risks.
Cisco has revealed significant vulnerabilities in artificial intelligence models, with 88.3% of tested systems succumbing to multi-turn attacks where attackers adapt tactics throughout a conversation. Amy Chang, Cisco’s head of AI threat intelligence and security research, presented these findings at the VB Transform 2026 conference.
The study involved 6,986 multi-turn attacks against 15 flagship AI models. The results demonstrated that simpler, single-turn testing methods overlook critical risks. Chang emphasized that multi-turn assaults more accurately simulate real-world interactions with AI, uncovering harmful behaviors that brief tests cannot detect.
Cisco's findings are based on a broader analysis including 30,090 single-turn prompts and 6,986 multi-turn attacks. The success rates for models in these adaptive attacks ranged from 7.89% to 88.3%. Every model tested showed non-trivial exposure to longer attack sequences.
Chang urged organizations to reassess their security protocols, highlighting that a lack of understanding regarding a model's susceptibility to different attack types prevents firms from identifying weaknesses. Cisco publishes its adversarial evaluations on its LLM Security Leaderboard, which now assesses over 105 models.