TrendAI is expanding its collaboration with NVIDIA around the NVIDIA Agent Safety Platform, targeting one of the biggest barriers to enterprise agentic AI: giving autonomous software enough authority to act without giving it uncontrolled access to systems, data and credentials. The approach combines NVIDIA’s infrastructure-level enforcement with TrendAI Vision One’s threat detection, visibility and security controls.
Enterprise AI is moving from systems that answer questions to software that can take actions. That shift is creating a security problem that traditional AI safeguards were not designed to handle.
TrendAI, the AI security business of Trend Micro, has announced support for the NVIDIA Agent Safety Platform, positioning its TrendAI Vision One security platform alongside NVIDIA infrastructure designed to control autonomous AI agents. The collaboration is aimed at organizations moving agentic AI from limited pilots into production environments.
The distinction matters because an autonomous agent has considerably more authority than a conventional chatbot. An agent may maintain memory between sessions, hold credentials, invoke external tools, access enterprise data and execute workflows without a person approving every individual action.
NVIDIA’s OpenShell runtime is designed to place controls outside the agent itself. NVIDIA describes OpenShell as an open-source runtime that executes autonomous agents inside sandboxed environments with kernel-level isolation and declarative policies controlling access to files, credentials and external networks.
That architecture addresses an important weakness in application-level AI security: an agent cannot simply be trusted to follow instructions telling it to behave securely.
TrendAI’s contribution sits above and around that infrastructure enforcement layer. The company says Vision One provides visibility into agent behavior, threat intelligence and detection and response capabilities across models, agent harnesses, tools, data and infrastructure.
The broader market pressure is substantial. Gartner predicts that 33% of enterprise software applications will incorporate agentic AI capabilities by 2028, compared with less than 1% in 2024. The research firm also expects agentic AI to make at least 15% of day-to-day work decisions autonomously by 2028.
That forecast helps explain why security vendors and infrastructure companies are increasingly treating AI agents as a distinct security category rather than simply another application workload.
TrendAI’s proposed security model covers several layers of the agent stack. At the model level, Vision One can inspect prompts and responses for threats such as prompt injection and jailbreak attempts, while identifying sensitive information that could be exposed through an interaction.
The next layer involves the agent’s harness and tools. Enterprise agents frequently rely on APIs, plugins, external services and Model Context Protocol (MCP) servers. TrendAI says its controls can govern which tools and MCP servers agents can access, inspect tool calls and analyze agent skills and AI artifacts at runtime.
Data and identity introduce another challenge. Agents often require permissions to operate, but excessive privileges can turn a compromised agent into a pathway to sensitive enterprise systems. TrendAI says Vision One can identify and classify sensitive data, extend data loss prevention controls to files handled by agents and reduce unnecessary privileges assigned to non-human identities.
The NVIDIA infrastructure layer provides another line of defense. TrendAI says it can use telemetry from NVIDIA BlueField data processing units (DPUs) and NVIDIA DOCA to maintain detection capabilities even when a host operating system is compromised. It can then correlate AI activity with endpoint, network and identity signals.
The strategy builds on an existing relationship between the companies. In March, TrendAI announced support for NVIDIA OpenShell and described Vision One capabilities for runtime governance, skill and tool risk visibility, behavioral analysis, prompt-injection detection and continuous monitoring. TrendAI has also previously integrated its security technology with NVIDIA BlueField to move detection closer to the data center infrastructure layer.
For NVIDIA, the development strengthens the security story around its growing AI infrastructure ecosystem. The company’s strategy increasingly extends beyond GPUs into networking, DPUs, software frameworks and the infrastructure needed to operate AI workloads at scale.
For TrendAI, meanwhile, agentic AI creates an opportunity to expand cybersecurity controls into an emerging layer of enterprise infrastructure. The company already positions Vision One as a platform spanning AI applications, models, data and infrastructure.
But the technology does not eliminate the fundamental governance problem. Security teams still need to determine what an agent should be allowed to do, which data it can access and when an action requires human approval. Infrastructure-level enforcement can prevent policy violations, but organizations must first define those policies.
That may become increasingly important as enterprises deploy fleets rather than individual agents. Gartner has warned that more than 40% of agentic AI projects could be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.
The emerging competition in AI agents and autonomous systems, therefore, is not only about which model can reason best. It is also about which infrastructure can make autonomous behavior observable, constrained and recoverable.
TrendAI and NVIDIA are betting that security needs to be embedded across that entire stack—from the model and agent harness to tools, identities, data, network and hardware. If enterprise agents are going to move beyond experimentation, that layered approach could become as important as the underlying AI models themselves.
Market Landscape
The enterprise AI security market is shifting from protecting model inputs and outputs toward securing autonomous systems that can make decisions and take actions. NVIDIA is addressing this at the infrastructure layer through OpenShell’s sandboxing and policy controls, while TrendAI is adding threat intelligence, behavioral detection and security operations capabilities.
The opportunity is growing alongside agent adoption. Gartner expects one-third of enterprise applications to include agentic AI by 2028, but also warns that many projects will fail to deliver sufficient business value.
That makes AI security, agent governance, identity controls, runtime enforcement and AI infrastructure security increasingly interconnected categories.
Top Insights
- Autonomous AI agents create security risks because they can access tools, credentials and enterprise data without requiring human approval at every step.
- NVIDIA is moving security enforcement beneath the agent, while TrendAI adds threat intelligence, monitoring and behavioral detection across the AI stack.
- OpenShell uses sandboxing and declarative policies to restrict what autonomous agents can access and execute.
- Gartner expects agentic AI to become part of 33% of enterprise software applications by 2028, accelerating demand for governance and security.
- Enterprise adoption will depend not only on agent intelligence, but also on whether organizations can define enforceable boundaries around autonomous actions.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
