Security Experts Warn Autonomous AI Agents Expand Enterprise Attack Surface

Security experts at Black Hat USA 2026 warned that the rise of autonomous AI agents is creating significant new vulnerabilities across enterprise environments. Nathan Hamiel of Kudelski Security noted that unlike traditional assistants, autonomous agents can execute code and call tools independently, introducing risks such as prompt injection and supply-chain attacks. He argued that runtime-generated code often evades static analysis, urging leaders to prioritize sandboxing and strict permission controls to limit the blast radius of agent behavior.

Addressing endpoint security, Lou Manousos of Ent stated that legacy tools are insufficient because they treat agents as standard processes and fail to monitor internal decision-making. He advocated for a security architecture capable of sub-second response times and real-time preventative defense to stop agents from misinterpreting user intent or accessing sensitive data.

Adding to these concerns, Tony Anscombe of ESET highlighted a growing threat where attackers embed malicious capabilities into downloadable AI skills to steal credentials and execute unauthorized code. Anscombe cautioned against blanket bans on AI, which he claimed drive employees toward shadow AI. He recommended a balanced governance approach starting with audits of employee usage. To address these threats, ESET released a free AI Skills Checker to analyze skill URLs for malicious activity in real time.