Cisco tested 15 top AI models with 6,986 multi-turn attacks at VB Transform 2026. Amy Chang, Cisco's AI threat intelligence lead, presented the findings. The results showed a significant vulnerability in current AI security. The tests were conducted to assess AI model defences.
The attackers adapted their approach across conversations, mimicking real-world scenarios. This method differs from single-turn testing, where the AI model is presented with a single prompt. By using multi-turn attacks, the researchers simulated a more realistic and dynamic threat environment. This approach revealed the AI models' weaknesses.
The tests revealed that 88.3% of the time, attackers successfully breached the AI models. The flagship models, considered to be among the most secure, were compromised by adapting the attack across the conversation. Chang's findings highlight the limitations of current AI security measures. The results indicate a significant gap in the defence mechanisms.
The research demonstrates that relying solely on single-turn testing is insufficient. AI models must be evaluated using more sophisticated and dynamic testing methods. The high success rate of multi-turn attacks underscores the need for improved security protocols. Chang's presentation at VB Transform 2026 emphasized the importance of addressing this vulnerability.
The consequences of these findings are far-reaching, with implications for organisations relying on AI models. As AI technology continues to evolve, the need for robust security measures becomes increasingly pressing. The vulnerability exposed by Cisco's research must be addressed to prevent potential breaches.
What is a multi-turn attack? A multi-turn attack involves adapting the attack across a conversation to bypass AI model defences. This approach simulates a realistic threat environment. It differs from single-turn testing.
How successful were the multi-turn attacks? The attacks were successful 88.3% of the time, highlighting a significant vulnerability in current AI security measures. The tests were conducted on 15 flagship AI models.
What are the implications of these findings? The findings indicate a need for improved security protocols and more sophisticated testing methods to protect AI models from adaptive attacks. Organisations relying on AI must address this vulnerability to prevent potential breaches.