Cisco’s AI security lead, Amy Chang, revealed at VB Transform 2026 that multi-turn attacks breached AI models 88.3% of the time in a study involving nearly 7,000 multi-turn attacks on 15 leading proprietary AI models. This highlights the insufficiency of single-turn testing, which fails to capture risks posed in extended conversational interactions. Cisco’s findings were supported by data showing over half of surveyed enterprises have faced AI agent security incidents or near-misses, underscoring the urgency for improved security frameworks. Industry giants like Palo Alto Networks and CrowdStrike are responding by investing heavily in identity and isolation security layers designed to safeguard AI agents. Chang and other experts emphasized the importance of multi-layered security involving scoped permissions, sandboxing, and continuous monitoring to mitigate risks. They also stressed the need to rethink security strategies from a fundamental level, moving beyond static code reviews to agentic security lifecycles where AI agents actively participate in enforcing protections. The discussion addressed challenges like intent detection versus probability assessment and the dynamic nature of AI agent vulnerabilities, urging enterprises to adopt continuous, multi-turn testing approaches to stay ahead of increasingly adaptive adversaries.
Back
