Eclypsium partners with Nvidia to lock down the hardware beneath AI agents

Eclypsium is teaming up with Nvidia to tackle a part of AI security most people never think about: the hardware and firmware sitting underneath the models.

The Portland-based infrastructure supply chain security company announced the partnership on October 29, 2026, via the Nvidia Blog. The goal is to build trust into AI infrastructure, especially cloud environments and the autonomous AI agents that increasingly run inside them.

What the partnership actually does

The deal combines Eclypsium’s AI Infrastructure Assurance platform with Nvidia’s security technologies. Two Nvidia pieces do the heavy lifting: its BlueField Data Processing Units (DPUs) and its Open Agent Safety Platform.

Eclypsium’s side is essentially a very thorough inspection service for servers. The platform takes inventory of hardware and firmware, then compares what it finds against known-good baselines. If something has changed without authorization, it gets flagged. The platform also surfaces known vulnerabilities in those components.

The scale of those blueprints matters. Eclypsium verifies more than 30 million known-good binaries when assessing hardware firmware posture.

Nvidia’s BlueField DPUs handle a different job. A DPU is a specialized processor that offloads networking, storage, and security tasks from the main CPU. In this partnership, the DPUs are used to create isolated trust domains for governing AI agents.

Layered together, the two companies describe the result as an end-to-end chain of trust. Each link, from the firmware up to agent policy execution, is meant to be checked rather than assumed.

Why firmware suddenly matters for AI

The most striking detail in the announcement comes from joint research the two firms conducted. It found that compromised infrastructure could trigger problematic in-memory model reasoning, even when the model hashes remained valid.

A model hash is a digital fingerprint used to confirm that a model file has not been altered. The research suggests that assumption has a blind spot: the model file can be pristine while the machine running it is not.

The partnership also responds to rising concern about supply chain attacks. Those are attacks that compromise components before they ever reach the data center, so the damage is baked in on arrival.

A relationship years in the making

This is not a first date. Eclypsium began supporting Nvidia hardware in May 2024. The relationship deepened on December 10, 2025, when Eclypsium was accepted into Nvidia’s Inception program. The new partnership builds on both of those earlier steps.

Eclypsium was founded in 2018 by security experts who previously worked at major firms, including Intel. The company serves cloud and AI data center operators. Its tools offer agentless monitoring, meaning operators do not need to install extra software on every machine to get visibility, along with integrity verification.

Funding has followed the thesis. Eclypsium has raised over $100 million in total, including a $25 million investment in March 2026. The company plans to use that capital to strengthen security across AI data centers.

What this means for AI infrastructure

The biggest takeaway is a shift in what “AI security” means. This partnership pushes the frame downward. If a compromised server can warp model reasoning without breaking a hash, then model-level safeguards alone cannot guarantee trustworthy outputs.

For Nvidia, the move fits a broader pattern of selling not just compute but the surrounding trust and governance layers. Using BlueField DPUs as an enforcement point gives its hardware a security role beyond raw performance.

For Eclypsium, the partnership offers distribution and credibility inside the ecosystem that dominates AI hardware. Being plugged directly into Nvidia’s agent safety tooling is a meaningful position for a company of its size.