Systems Security Foundations for Agentic Computing

Agentic AI systems face unique security vulnerabilities because they constantly interact with external tools and servers (web browsers, compilers, etc.), creating attack surfaces that traditional standalone AI models don't have—making them susceptible to threats like prompt injection attacks that can lead to data exfiltration.

Current AI safety research lacks cybersecurity rigor: Most existing work on agentic AI safety focuses on the AI model itself rather than examining end-to-end security properties of entire systems, leaving important gaps that decades of established cybersecurity principles could help address.

Real-world attacks on agentic systems are already happening: The paper documents 11 case studies of actual attacks on agentic systems, demonstrating that these security concerns are not theoretical but represent immediate, practical threats that organizations deploying these systems need to address.

Traditional security principles need adaptation for AI agents: The paper identifies that applying classic cybersecurity research to agentic AI reveals new research challenges specific to this domain, suggesting that neither pure AI safety approaches nor conventional cybersecurity alone is sufficient.

Bridging AI and cybersecurity expertise is critical: The paper emphasizes the need for collaboration between AI practitioners and security researchers to develop comprehensive safeguards, as the intersection of these fields requires both perspectives to adequately protect agentic systems.