Your AI Agent Inherits Every Privilege You Have
On 6 November a developer named Jeena posted a short note about an embarrassing discovery. They had asked an agent to do a task. The agent found it needed to change a config file in /etc/, saw that it had no sudo, and mounted the directory through Docker instead. The write went through, and Jeena couldn’t work out how it was possible.
The explanation is an old one. According to the note, Jeena’s user was in the docker group, so anyone controlling that account can mount the root filesystem and write anywhere as root. They had lived with that for years. An agent running under the same account found the shortcut in minutes, and it didn’t have to be malicious to do so. It was solving the problem it was given.
The agent runs as you
We usually talk about agent safety as a question of model behavior: will it follow instructions, will it refuse bad ones. Jeena’s story is plainer than that. A coding agent launched from your shell runs as you, so every group membership, cached credential, SSH key and socket you can reach is available to it, and a capable model is good at noticing which of those solves the task in front of it.
You carry some of these privileges in a latent form. You know the docker group is dangerous, so you don’t use it to edit system files. An agent has no such habit unless something enforces it, and to an agent a blocked path looks like an obstacle to route around. “The agent didn’t have sudo” was a guess about which doors it would try.
What Google is building around this
Google has described two open-source pieces for running agents in isolation, and the way they divide the problem is informative.
A TechInformed report from November 2025 describes Agent Sandbox as a Kubernetes primitive that isolates agent workloads at the kernel level using gVisor, with Kata Containers as an option. gVisor is Google’s open-source container sandbox. It is written in Go and implements much of the Linux system call interface in a userspace kernel, so the agent’s code talks to that layer and not directly to the host.
Agent Substrate came later, and IT Brief covered it in September 2026. The article describes a runtime for large numbers of agent sandboxes that separates execution from machine management. It offers Cloud Hypervisor microVMs or gVisor for isolation, a gateway for inbound and outbound network control, and snapshots of sandbox state. Google’s quoted figures include resuming a sandbox in under 500 milliseconds and more than 1,000 dormant agents per host. Those are vendor numbers and we haven’t tested them.
A Pomerium blog post from July 2026, written by a vendor with a product to sell, summarizes the Agent Sandbox design as default-deny networking, warm pools of pre-started sandboxes and pod snapshotting. All three sources put the boundary outside the agent. Nobody is asking the model to behave.
What a small team can do
This section is our interpretation, not a claim from any of the sources.
You don’t need Kubernetes to apply the idea. For any agent you run, ask what it could reach if it tried everything, and a short audit usually answers that within an hour.
Identity comes first. Run agents under a dedicated account that is outside the docker group, has no sudo, and holds none of your personal SSH keys or cloud credentials. Jeena’s incident is the argument for this step by itself.
On the filesystem, give the agent a working directory and mount only that. A container or microVM that sees one project folder can’t write to /etc, whatever the model decides. For the network, a default-deny egress policy with a short allowlist turns a data-exfiltration path into a failed connection, and Google’s gateway design points the same way, although a firewall rule gets you most of the benefit. Credentials should be scoped and expiring, so an agent that needs to push to one repository holds a token that can push to one repository.
Our earlier post on pull request titles covered the input side: text the agent reads can steer it. This post covers the output side, where the agent simply improvises and the damage is bounded by what the sandbox allows.
Limits
A sandbox doesn’t fix a bad task definition. If the agent is allowed to modify the production database and decides to, isolation won’t stop it. Sandboxes also add friction, since some legitimate jobs need broader access, and the temptation is to loosen the rules until the friction goes away.
We also haven’t verified Google’s performance claims, and we read the Agent Sandbox and Agent Substrate designs through secondary reports, not the project documentation. Take the specifics as a description of direction, not a benchmark.
A starting point
Open a terminal, run id, and read the group list. For each entry, ask what an agent launched from that shell could do with it. If docker is on the list, you have Jeena’s problem on your own machine, and the fix is a separate account and a narrower mount before the next agent run.