AI Agent Security Solutions

Cisco DefenseClaw runs coding agents in a sandbox

Cisco described DefenseClaw with NVIDIA OpenShell on 21 October 2026. Per Cisco, DefenseClaw checks every prompt, tool call and tool result inside the sandbox, decides web destinations by host name, hides real provider credentials from the agent, and records each block and approval for Splunk or OpenTelemetry backends.

Facts read on .

Cisco published a blog on 21 October 2026 on protecting enterprise coding agents with DefenseClaw and NVIDIA OpenShell. The premise is that a coding agent acts with the full access rights of the developer who started it, so a sandbox should limit it to what the task needs.

Per Cisco, OpenShell runs each agent in a sandbox where per-agent network egress policy acts as micro-segmentation, and where the agent sees placeholders instead of real provider credentials. DefenseClaw checks each prompt, tool call and tool result using its own rules and Cisco AI Defense, and decides each web destination by host name without decrypting traffic.

Administrators can set user- or group-based policy for which coding agents run on managed endpoints and which run only in a sandbox. Cisco says every block, approval and decision is recorded with the agent, sandbox and policy behind it and can go to Splunk or any OpenTelemetry backend. These are Cisco's statements and were not tested by this site.

Sources