Agentic AI security requires human oversight, Dash Security says

Agentic AI security, a new approach that inverts traditional threat models by granting autonomous software standing access to critical systems, was discussed at the Fal.Con event. This model requires a shift in how security is approached, as these AI agents are invited into the business and given delegated authority, based on information published by Silicon Angle.

Traditional security assumes external attackers, but agentic AI operates with delegated authority, making it appear legitimate until it deviates. Amol Mathur, co-founder and chief product officer of Dash Security, explained that AI agents are non-deterministic and mix data, instructions, and input, necessitating runtime governance.

Dash Security focuses on providing visibility into AI tool usage within organizations and implementing "guardrails" to enable AI adoption without hindering operations. Mathur highlighted a case where precise understanding of an agent's role reduced false positives by 95%, allowing for automated actions. Dash integrates with CrowdStrike, feeding endpoint telemetry into its SIEM platform. The company emphasizes understanding both the evolution of generative and agentic AI and how customers adopt these technologies.

Source: Silicon Angle