Multi-turn attacks broke AI models 88% of the time — single-turn testing missed it, Cisco AI security lead warns at VB Transform 2026

Cisco’s Amy Chang said tests of 15 flagship AI models found multi-turn attacks succeeded as often as 88.3%, while single-turn red-teaming missed key risks. A VentureBeat survey found 54% of enterprises have already seen an agent security incident or near-miss, yet only about a third use scoped identities or sandboxing for agents.

The findings suggest AI security programs must move beyond prompt-level testing toward continuous adversarial evaluation across conversations and workflows. Big security acquisitions around identity and isolation show vendors are racing to secure the agentic AI layer before enterprises fully harden it themselves.