What happened

Red Hat’s vice-president and general manager of Red Hat AI, Joe Fernandes, told SiliconANGLE that agents differ from conventional software because they are non-deterministic and act to achieve a goal without a fixed sequence of steps. Senior distinguished engineer Jessica Forrester added that Red Hat is building layered guardrails, sandboxing, identity management, access controls and enterprise-scale agent management, to contain autonomous agents.

The firm has spent the past year creating a sandbox for OpenClaw agents and integrating the open-source OpenShell runtime, launched by Nvidia in March, into its security stack. OpenShell can run agents in Kata Containers with their own kernels, inside full virtual machines on OpenShift Virtualization, and behind network egress controls. Red Hat also masks credentials that agents receive, swapping them on exit so that no real keys leave the sandbox.

Why it matters

Traditional security tools rely on the premise that software behaves exactly as coded. Non-deterministic agents break that premise by selecting any path that satisfies a high-level objective. This expands the potential damage surface: every tool or credential an agent touches becomes a possible point of compromise. By masking keys and assigning each agent a narrow identity, Red Hat aims to limit what an agent can do and, crucially, what it cannot do.

Red Hat’s partnership with OpenAI on the OpenClaw Enterprise project shows the scale of the challenge, managing thousands of personal agents in production. The company says the current preview of OpenShell marks a stable milestone that could give enterprises confidence to run autonomous agents in production environments.

What this means for enterprise IT teams

Enterprises must rethink security policies to accommodate agents that do not follow a predetermined code path. Controls such as sandboxing, credential masking and per-agent identities become essential. IT teams should evaluate runtimes that provide kernel-level isolation and enforce strict egress filtering. Monitoring tools need to capture agent behaviour in real time to detect unexpected actions before they cause harm.

Adopting a layered approach, combining container or virtual-machine isolation, masked credentials, and fine-grained identity and access management, will help reduce the risk of an agent inadvertently exposing sensitive data or executing malicious commands. Teams should also prepare for the operational overhead of managing large numbers of agents, including lifecycle management and compliance reporting.

In the words of Joe Fernandes, vice president and general manager of Red Hat AI: "Agents are different. They’re non-deterministic; you’re giving them a goal. You don’t know what actions they’re going to take to get to that goal."

The bottom line

Non-deterministic AI agents overturn the deterministic assumptions that underpin most security models. Red Hat’s sandboxing, masked key handling and per-agent identity strategy illustrate one path forward. Enterprises that wish to harness autonomous agents must adopt similar layered safeguards to protect critical systems and data.

Sources

  • siliconangle.com