Anthropic and Nvidia Unveil Zero-Trust AI Agent Architectures

Two new designs tackle credential exposure and the 'blast radius' of compromised AI agents

Apr. 10, 2026 at 10:26pm

Got story updates? Submit your updates here. ›

A highly detailed, glowing 3D illustration of various cybernetic hardware components, including circuit boards, processors, and cables, all illuminated by vibrant neon cyan and magenta lights, conceptually representing the complex digital infrastructure underlying AI agent security.As enterprises race to deploy AI agents, new zero-trust architectures aim to secure the digital foundations and limit the 'blast radius' of compromised agents.Denver Today

Anthropic and Nvidia have released the first zero-trust AI agent architectures, addressing the growing security concerns around AI agents that hold sensitive credentials and can execute untrusted code. The two companies have taken different approaches, with Anthropic separating the 'brain' from the 'hands' of the agent and Nvidia locking down the sandbox and monitoring everything inside it. The gap between their designs reveals where the real risk lies for enterprise security teams.

Why it matters

As AI agents become more prevalent in enterprises, with 79% of organizations already using them, the security risks are becoming increasingly clear. The default monolithic agent pattern leaves credentials exposed and the entire container at risk if compromised. The new architectures from Anthropic and Nvidia aim to reduce this 'blast radius' and better isolate sensitive credentials from untrusted code execution.

The details

Anthropic's Managed Agents split the agent into three components that do not trust each other: a brain, disposable execution containers ('hands'), and a persistent session log. Credentials never enter the sandbox, instead being fetched by a proxy. Nvidia's NemoClaw, on the other hand, wraps the entire agent in multiple security layers, including intent verification to monitor and approve every action. The key difference is in the proximity of credentials to the execution environment - Anthropic removes them entirely, while Nvidia gates them through policy.

  • Anthropic launched Managed Agents in public beta on April 8, 2026.
  • Nvidia released NemoClaw in early preview on March 16, 2026.

The players

Anthropic

An American autonomous driving company and is a subsidiary of Alphabet Inc., Google's parent company.

Nvidia

A technology company known for its graphics processing units (GPUs) and AI computing platforms.

Matt Caulfield

VP of Product for Identity and Duo at Cisco.

Elia Zaitsev

CTO at CrowdStrike.

George Kurtz

CEO of CrowdStrike.

Got photos? Submit your photos here. ›

What they’re saying

“While the concept of zero trust is good, we need to take it a step further. It's not just about authenticating once and then letting the agent run wild. It's about continuously verifying and scrutinizing every single action the agent's trying to take, because at any moment, that agent can go rogue.”

— Matt Caulfield, VP of Product for Identity and Duo at Cisco

“A lot of what securing agents look like would be very similar to what it looks like to secure highly privileged users. They have identities, they have access to underlying systems, they reason, they take action. There's rarely going to be one single solution that is the silver bullet. It's a defense in depth strategy.”

— Elia Zaitsev, CTO at CrowdStrike

What’s next

The judge in the case will decide on Tuesday whether or not to allow Walker Reed Quinn out on bail.

The takeaway

This case highlights growing concerns in the community about repeat offenders released on bail, raising questions about bail reform, public safety on SF streets, and if any special laws to govern autonomous vehicles in residential and commercial areas.