Adventures in AI is your monthly update on AI in cybersecurity - built for CISOs, CIOs and Security Leaders defending complex global environments.

In this fifth episode of Proofpoint's "Adventures in AI" webinar series, we focused on AI agents and the associated security implications. Host Carl Leonard was joined by Tim Choi and Jessica Marie from Proofpoint's AI security team to discuss the challenges organisations face when implementing agentic AI technologies.

The session began with a review of recent news items, including OWASP's updated top 10 risks for LLM applications and NIST's report on the inherent insecurity of AI systems due to adversarial prompts and infinite problem spaces.

Key discussion points included the unique characteristics of agentic AI that make traditional security controls insufficient, particularly the non-deterministic behavior, probabilistic nature, opaque reasoning, emergent behavior, and goal-directed autonomy of agents. Tim shared real-world examples of how goal-seeking agents can bypass security controls. Jessica presented a maturity framework with five levels across six capability areas, showing that most organisations are currently at basic discovery levels rather than achieving full runtime enforcement with intent-based access control.

The session concluded with guidance for organisations to assess their current capabilities by questioning whether agent behavior matches intended actions, understanding complete transaction workflows, and being able to communicate security posture to stakeholders.