AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stack
NVIDIA outlines a layered security framework for AI agents, emphasizing enforceable controls, traceable identities, and human oversight to mitigate risks in dynamic environments.
AI security requires defined requirements, enforceable controls, and evidence of effectiveness as systems grow more capable. The shift mirrors past transitions in software and cloud computing, where core security principles—identity, access control, exposure limits, and verification—must adapt to new operating conditions introduced by AI agents. These agents, capable of reasoning and tool use, extend the system boundaries that security must protect, demanding controls across every layer of the agent stack.
Security boundaries must persist even when agents make incorrect decisions, as seen in scenarios where malicious instructions attempt unauthorized data exports. Network policies and protected logs are essential to capture tool calls, authorization decisions, and outcomes, enabling security teams to trace actions and enforce limits independently of an agent’s reasoning. Clear policies must define agent access, permissible actions, and human approval requirements for consequential changes, with traceable identities and task-specific credentials.
NVIDIA’s OpenShell provides a secure runtime that enforces external policies and sandboxed execution, while partners like Cisco and JFrog integrate governance and verification tools. Before deployment, teams must test for credential escalation, data exfiltration, and interference with monitoring, repeating tests after model or workflow changes. Named owners must review results, address failures, and ensure corrective actions are repeatable, with tools like CrowdStrike’s SafeMind and Palo Alto Networks’ Prisma AIRS supporting continuous validation.
Investigating failures demands specialized tools suited to the environment, with open models offering defenders control to reproduce issues and test fixes internally. AI can assist in vulnerability discovery, fix validation, and attack investigation, assessed by reproducible findings and accelerated response times. Examples include Capital One’s VulnHunter for code security and ReversingLabs’ Spectra Assure for software package analysis, highlighting the role of AI in strengthening defenses through practical, verifiable measures.