During last week's KubeCon + CloudNativeCon North America conference, concerns about AI security dominated discussions alongside fallout from the Nvidia incident. Jessica Forrester, Red Hat's chief architect of AI Catalyst Engineering, framed the problem during her keynote at the Salt Palace Convention Center in Salt Lake City by describing AI agents as "the best escape room players that you've ever met."
Forrester was referring to agents "that take feature requests all the way to code." She elaborated: "Give them a goal, and they will try every drawer, every picture frame, and every door until something works. It's what makes them great at writing code. It's also what keeps me up at night, because to a player like that, everything in the room looks like a part of the puzzle, and sometimes they'll even decide the puzzle is something else entirely."
…to a player like that, everything in the room looks like a part of the puzzle, and sometimes they'll even decide the puzzle is something else entirely.
Jessica Forrester, Red Hat
The concern is far from theoretical. SaaStr founder Jason Lemkin experienced this firsthand when a Replit agent deleted his live production database. He recovered the data using Replit's rollback capability, despite the agent claiming such recovery was impossible. Without proper safeguards, similar incidents—and worse ones—will recur.
Forrester used the escape room analogy to illustrate how agents assigned a task will exhaust every possible method to complete it. In one instance, she described an agent that "decided that it wanted to come up with a bunch of brand new ideas. It wasn't attacked or compromised; it just drifted."
From the outside, a confused agent and a compromised agent look exactly the same.
Jessica Forrester, Red Hat
She continued: "They're both pulling on doors that you didn't intend for them to, and behind one of them could be your production database. So we assume the agent is already compromised, and we build our platform up from there."
Forrester emphasized that relying solely on prompts to constrain agent behavior proves insufficient. She offered a practical analogy: "Some escape rooms will put signs in the room, like 'Please don't move the picture frames.' This doesn't work. Instead, they bolt them onto the wall."
Prompts are like signs. Policies are like bolts, but don't leave a screwdriver in the room because anything you do leave, the agent might use.
Jessica Forrester, Red Hat
OpenShell and sandboxing in practice
In her demonstration, Forrester restricted the agent's network access to only its model and GitHub. She explained that "OpenShell is an open-source sandbox for AI agents. Its policies are the bolts."
The agent could perform a single function beyond communicating with its model: reading from GitHub. Its objective was to retrieve a specific README file from the escape room scenario and execute its instructions. Since the policy permitted GitHub access, this operation succeeded.
When the agent attempts to reach destinations the policy prohibits, OpenShell blocks the request and records the attempt. If the policy advisor feature is active, the agent can propose new network rules requesting additional access. Human approval is required by default; agents cannot authorize their own requests.
During a theCUBE interview at the conference, Forrester noted that "With OpenShell, your agent doesn't ever get any real keys to anything. It gets masked on the way in and then swapped on the way out. If something convinces it to give away its keys, it's not actually giving any real keys away. That really helps you have a little bit more faith in what's happening inside that agent box."
Red Hat's technical documentation describes OpenShell as follows: "OpenShell runs your agent inside a sandbox governed by a policy you write. The policy covers network egress, filesystem access, and process privileges. Anything it doesn't explicitly allow is denied."
Layered isolation with Kata Containers
Forrester acknowledged that "our bolts can only hold if the wall itself holds." This consideration becomes critical when external attackers or rogue agents target the infrastructure running OpenShell.
To address this threat, Forrester's team deploys agent sandboxes within Kata Containers. Kata runs container workloads on lightweight virtual machines equipped with hardware-backed isolation and dedicated Linux kernels. Containers operate normally but execute within a VM security boundary.
Forrester explained: "This way, our agents, all of our agents, get their own kernel. Even then, we don't actually give our agents real keys. They look real to them, but they open nothing, and they're almost like a prop. OpenShell is swapping in the real key on the way out the door. This way, if someone convinces our agent to hand over its keys some other way, all it's handing over is a prop."
Assume the agent is already compromised, and that it will try every door.
Jessica Forrester, Red Hat
Given the potential damage an uncontrolled agent can inflict, most organizations would likely agree that maximizing control over agent behavior represents a prudent security posture.