AI agents' cybersecurity risks addressed in training

Summary

A Turkish cybersecurity association organized a hands-on workshop where participants deployed an AI security agent and tested it against prompt injection attacks. The association's chair argues that agentic AI's ability to autonomously plan and execute actions necessitates strict permission limits and human-in-the-loop oversight mechanisms.

Why it matters

Agentic AI systems capable of independent planning, tool invocation, and decision-making create a qualitatively different threat surface than traditional generative models, as erroneous decisions can directly impact live systems. The event highlights a growing gap: organizations are deploying these agents faster than they are establishing the necessary governance and security controls. Positioning 'safe AI development' as a distinct field of expertise signals that cybersecurity curricula and hiring processes must rapidly adapt.

Highlights

• Participants deployed a security agent in a live environment and demonstrated the impact of prompt injection on agent behavior.
• Association chair: "The more capabilities we give to AI, the more we are forced to manage those permissions."
• Uslu argues that simply using or building AI is no longer sufficient; developing *safe* AI systems is becoming a separate and critical skill set.

Critical take

The text reads like a press release for a single association event; it lacks independent expert quotes, concrete incidents, or data points. The 'Hackviser accessible prize competition' framework dilutes what is otherwise a serious topic. A critical question—how many organizations are actually implementing the permission limits and human-approval gates advocated by Uslu?—is completely ignored.