AI Security Must Be Treated as an Engineering Problem: Emphasizing Control and Verification Across All Layers of the Agent Stack

AI security is not merely a policy but an engineering problem characterized by defined requirements, enforceable controls, clear accountability, and evidence proving the effectiveness of protective measures. As AI's reasoning and tool-use capabilities improve, existing security principles must be adapted to new operational conditions.

Security is required across the entire Agent Stack, comprising the model, harness, and runtime environment. To maintain security boundaries even if an Agent makes erroneous decisions, file, network, and process access must be restricted at the environment level independently of the Agent's reasoning. Each Agent must have traceable identity and credentials scoped to its tasks, with human approval required for critical operations or privilege changes.

Verification and Tool Utilization

The provenance and integrity of tools, skills, and dependencies must be verified, and protected logs of tool calls and permission decisions must be retained to support incident reconstruction and isolation procedures. NVIDIA OpenShell is an open-source security runtime that enforces policies outside the Agent's scope, managing access to data and system resources. Additionally, Cisco DefenseClaw and JFrog enforce access policies by adding governance layers and scanning Agent skills.

Before deployment, the effectiveness of controls must be proven to block attempts to obtain credentials exceeding the Agent's scope or to transmit unauthorized data. CrowdStrike SafeMind performs iterative attack simulations, while Palo Alto Networks Prisma AIRS conducts continuous red teaming during model and app changes to strengthen defenses. Closed models enable managed services, while open models allow component inspection and self-hosted infrastructure operations, playing complementary roles in defense.