NSS Labs outlines runtime guardrails and an enterprise AI security evaluation framework
NSS Labs published two white papers that frame enterprise AI security as an evaluation problem focused on runtime guardrails, governance, and evidence. The update matters to security and risk leaders selecting controls for production AI use cases.
Research overview
The March series consists of two parts: a paper on why protecting only the model is insufficient and a paper that provides questions for evaluating enterprise AI security. The work is organized around how risk emerges from components around an AI system rather than within the model alone.
NSS Labs presents the evaluation effort for CISOs, enterprise buyers, GRC leaders, and security architects. It aims to support vendor assessment under conditions intended to reflect enterprise deployments.
Key findings
The first paper argues that the primary risk in real deployments is found in surrounding systems, including data sources accessed, tools invoked, permissions inherited, policies applied, and the visibility available into system behavior. It ties those themes to input integrity, output risk, resilience, policy governance, agentic behavior, observability, and GRC.
The series identifies runtime guardrails as controls that operate outside the model. These guardrails are described as enforcing policy, constraining access, mediating tool use, reducing data leakage, and producing evidence needed by security and governance teams when issues occur.
Evaluation framework and governance requirements
The second paper turns the architecture and governance framing into an evaluation framework for comparing AI security controls for production environments. It emphasizes questions buyers should ask, warning signs to watch for, and criteria used during vendor comparisons.
The series also states that AI security cannot be separated from governance. It calls for controls that can be explained, tested, monitored, tuned, and audited, with traceability for decisions, policy triggers, data access, tool invocation, and constrained authority.
Market and testing direction
NSS Labs describes a gap between the speed of AI security developments and the availability of rigorous evaluation standards. It says the gap can make it harder to distinguish engineering from vendor narratives and harder for vendors with controls to validate claims through disciplined testing.
NSS Labs says it has begun testing AI Protection Systems (AIPS). It positions the two papers as a starting point for evaluating AI security products by clarifying what matters, what questions to ask, and what it describes as good.
NSS Labs’s two-paper series provides an enterprise AI security perspective centered on runtime guardrails, governance, observability, and auditable evidence, along with an evaluation approach aimed at production environments. This “Blog Signals” brief is a fact-based summary of the vendor blog.