article --summary
The summary
This article shares the views of Jessica Forrester, Red Hat's chief architect of AI Catalyst Engineering, on the security risks posed by autonomous AI agents. Her core idea is that an agent given a goal will try every option available, so it can drift off course or be manipulated in ways that look the same from the outside. Readers will see why she recommends assuming agents are already compromised and designing systems around that.
The piece walks through how Red Hat applies this thinking in practice. Its demo restricts an agent's network access with OpenShell, which blocks unapproved destinations, logs attempts, and requires human approval for new access requests. The agent also works with masked credentials that are swapped for real ones only on the way out, and its sandboxes run in Kata Containers so each agent gets its own kernel inside a lightweight VM.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.