NVIDIA launches agent safety platform with TrendAI

NVIDIA has launched the NVIDIA Agent Safety Platform, with support from TrendAI, for enterprises moving autonomous AI agents into production.

TrendAI is extending the platform with threat intelligence, monitoring and security controls across AI systems as companies grapple with the risks of agents that can access data, credentials and operational tools without step-by-step human approval.

AI agents are drawing closer scrutiny because they do more than generate text or analyse information. In enterprise settings, they can use software tools, retain context across sessions and carry out tasks inside production systems. That raises the stakes if a model is manipulated or a connected system is exposed.

Security for these systems should not focus on a single point, the companies said. Their approach is based on the view that risk sits across the model, the orchestration layer that runs the agent, and the tools and data sources the agent can reach.

According to NVIDIA, the platform uses OpenShell to enforce policy outside the agent's execution environment. It also uses NVIDIA BlueField-4 DPUs and NVIDIA DOCA to add a hardware-based security layer beneath the host system.

TrendAI said its Vision One platform adds oversight on top of that infrastructure layer, including threat intelligence, visibility into agent activity, and detection and response functions designed to help security teams define and monitor the rules under which agents operate.

Layered controls

Those controls span several parts of an AI deployment. Prompt and response inspection can be used to block prompt injection and jailbreak attempts, while redacting sensitive data before it is exposed, TrendAI said.

Security teams can also govern which tools and model context protocol servers agents are allowed to access, inspect each tool call and analyse how AI skills and other artefacts behave during runtime, according to the company. In practice, that means an organisation can limit what an agent touches and monitor what it does once deployed.

On the data side, TrendAI said its software can find and classify sensitive information used by AI systems, prevent agents from reaching restricted data, and apply file security and data loss prevention to material agents read or write. It added that the system can reduce excessive access rights for agents and other non-human identities.

Another element is host-independent monitoring. Telemetry from NVIDIA BlueField DPUs and DOCA allows detection to remain reliable even if the host operating system is compromised, TrendAI said, while correlating AI-related signals with endpoint, network and identity data to spot attempts at lateral movement.

Production hurdle

The announcement comes as businesses test whether agentic AI can move beyond controlled pilots into wider operational use. One of the main constraints has been the difficulty of governing systems that can take actions autonomously while interacting with sensitive corporate systems.

The issue has broader relevance in Australia, where concerns about AI agents reaching important systems and credentials have sharpened debate over how much autonomy organisations should allow software agents to have, and what controls should surround them.

Rachel Jin, Chief Platform and Business Officer, Head of TrendAI, outlined the company's view of that challenge in a separate commentary on agent security. "A well-intentioned agent can still do real damage. An agent is only as safe as the model it calls, the harness that runs it, and the tools and data it reaches. Attackers will target whichever layer is weakest. You don't choose between them; you see, manage, monitor and govern each layer," Jin said.

In the formal launch announcement, Jin described the partnership in broader terms. "Agentic AI is the most significant shift in enterprise technology in a generation, and security needs and guardrails are in place to keep the agents aligned with the intentions of the organisation," she said.

She added that a single control point would not be enough as these systems become more widely deployed. "Security can't sit in one place. It has to cover the model, the harness, and every tool and data source an agent touches. NVIDIA is building the boundary that determines what an agent can do. TrendAI brings the threat intelligence and oversight that tell security teams what that boundary should be, and what is happening inside it. Together, we give organisations a practical path to scale agentic AI securely," Jin said.