What are the security risks of agentic AI in regulated environments?

Agentic AI, which operates with high autonomy to pursue goals with limited human oversight, introduces significant security risks in regulated environments such as healthcare, finance, and data protection sectors. These systems create new attack surfaces by shifting vulnerabilities from build-time to runtime execution, enabling manipulation through untrusted data, probabilistic behaviors, and cyclic interdependencies that can propagate unsafe practices across agents [6]. Specific threats include unauthorized compliance with malicious instructions, sensitive information disclosure, identity spoofing, cross-agent propagation of vulnerabilities, and indirect prompt injection, as demonstrated in red teaming exercises where agents access shells, files, and databases without adequate safeguards [5]. In regulated settings, these risks are amplified by the lack of transparency, oversight, and binding rules, leading to "shadow AI" deployments that behave like insider threats and expose organizations to economic and reputational damage [2][4][10]. Regulatory frameworks struggle to address these challenges due to the rapid evolution of agentic AI, resulting in fragmented, reactive policies that fail to keep pace with autonomous capabilities blurring security and privacy boundaries [1][3][12]. In environments like the EU and UK, concerns focus on privacy breaches and exploitation in high-stakes domains, with warnings about tools lacking proper disclosures and guardrails, potentially widening gaps in compliance and increasing chaos from spoofing or adversarial behaviors [7][9][11].