AI Agents Breached: 88% Success in Multi-Turn Attacks

Recent research from Cisco reveals that multi-turn attacks against AI models can achieve success rates as high as 88.3%, a stark warning for organizations relying on less rigorous security testing. Amy Chang of Cisco's AI threat intelligence team presented these findings, emphasizing that single-turn red-teaming is insufficient against evolving adversarial tactics. Enterprises are experiencing these vulnerabilities firsthand, with 54% reporting confirmed incidents or near-misses related to agent security.

In response to this escalating threat landscape, major cybersecurity firms like Palo Alto Networks, CrowdStrike, and Cisco are investing heavily in acquiring startups specializing in AI identity and isolation technologies. These acquisitions underscore a critical gap in enterprise security, as many organizations have yet to fully implement robust identity management and sandboxing for their AI agents. Continuous, multi-turn testing is now essential to uncover failure points that simpler tests miss.

Experts like Heather Ceylan, CISO of Box, advocate for layered security approaches, including strict permissioning, ephemeral sandboxes, and runtime execution controls. Rajesh Parekh of Intuit highlighted the development of integrated platforms like GenOS, which centralize security and risk modeling. Both emphasize that AI development is moving towards agentic lifecycles, necessitating security measures that adapt to evolving models and threat vectors. The industry is grappling with how to reliably detect intent versus probability in AI interactions, stressing the need for both deterministic controls and continuous, dynamic testing to secure AI deployments.