Hope Is Not a Security Strategy: Why Secure-by-Default Beats Hardening (tuananh.net)

A recent article argues that the AI/ML community must shift from a "hope-based" security strategy to a "secure-by-default" approach, particularly as AI agents become more integral to enterprise operations. The author highlights the limitations of current security practices, which often rely on hardening and monitoring shared kernel infrastructures, making them vulnerable to exploits. With the rise of AI agents that exhibit underdeterministic behavior, traditional methods fall short, as they assume predictable actions and outcomes, leaving systems exposed to potential risks. To address these vulnerabilities, the piece advocates for using technologies like gVisor, Firecracker, and microVMs to isolate workloads and treat them as untrusted from the start. By designing systems with strong isolation and sandboxing protocols, organizations can better manage the unpredictability of AI agents, ensuring that security is not an afterthought but a foundational element. The article stresses that a secure-by-default approach should be accessible and easy to implement, moving away from complex setups that seem to favor larger enterprises. As the AI landscape evolves, adopting this security philosophy could redefine best practices and enhance trustworthiness within the sector.