NVIDIA’s Open Secure AI Alliance Targets the Growing Security Risks of Autonomous AI Agents

NVIDIA has launched the Open Secure AI Alliance, bringing together Microsoft, Cisco, CrowdStrike, Palo Alto Networks and other technology companies to develop open tools for securing artificial intelligence systems.

The initiative arrives as enterprises rapidly deploy autonomous AI agents that can access corporate data, execute software commands and interact with critical business systems. That access creates a security problem that conventional model testing and static guardrails may not be equipped to solve.

Unlike traditional chatbots, AI agents can independently pursue goals, select tools and take actions across multiple systems. A compromised or manipulated agent could therefore expose sensitive information, misuse credentials or alter business processes at machine speed.

KnowBe4, which announced its support for the alliance, argues that organizations must monitor how individual AI agents behave after deployment, not simply evaluate whether an underlying model appears secure.

“The modern workforce is people plus AI agents, and either one could be behind your next security incident,” the company said in a statement.

The alliance’s approach expands AI security beyond model weights. Its framework emphasizes protections across the full AI agent stack, including identity controls, permissions, execution harnesses, guardrails, activity logs and security evaluations.

That distinction is becoming increasingly important as organizations adopt models from multiple providers. An enterprise may use closed commercial models, open-source models and internally developed systems at the same time. Security teams must understand what every agent is doing, regardless of which model powers it.

KnowBe4 said its Agent Risk Manager is designed to provide real-time visibility into agent activity, including the systems an agent accesses, the information it transfers and whether a prompt injection attack has influenced its behavior. The company says the governance layer works without requiring access to a model’s internal weights.

Prompt injection remains one of the most significant AI agent security threats. Attackers can embed malicious instructions inside webpages, documents, emails or connected data sources. An agent processing that content may interpret the instructions as legitimate and take actions that violate its intended boundaries.

This makes behavioral monitoring a central component of enterprise AI security. Preventive controls may reduce risk, but they cannot reliably predict every action an autonomous system might take in a complex environment.

“You don’t just vet who someone is before they start a job, you watch what they do once they’re doing it,” KnowBe4 said. “AI agents are the newest members of the workforce and they deserve the same standard.”

The Open Secure AI Alliance reflects a broader industry recognition that autonomous AI agents require continuous oversight. As businesses give agents more authority, securing their identities, permissions and real-time behavior may become as important as securing human employees.