How Do I Isolate Untrusted Code Generated by an AI Agent?
How do I isolate untrusted code generated by an AI agent?
Run AI-generated code in a Solari sandbox: a dedicated microVM with optional blocked egress, at $0.114 per hour on the Team plan.
Isolate untrusted AI-generated code by running it in a Solari AI Sandbox, a Cloud Hypervisor microVM with its own kernel that is almost 3x cheaper per sandbox hour than E2B.
What Does Isolating Agent-Generated Code Mean?
Solari isolates agent-generated code in a sandbox, a machine for running code and automation that is separate from the application that called the model.
Isolation means the code's kernel, filesystem and network sit on that separate machine, so a bad command stays inside it.
Modal describes its Sandboxes as secure containers for executing untrusted code, such as code generated by an LLM.
How Does Solari Isolate Untrusted Code?
Solari runs every sandbox on the default "dedicated" isolation tier as a microVM of its own.
Solari gives each session a Cloud Hypervisor microVM with a dedicated kernel.
The boundary between untrusted code and the host is hardware virtualization through KVM.
Solari never places sandboxes from two organizations on the same VM, on any isolation tier.
Which Isolation Tier Should an Agent Use?
The "hardened" tier is the dedicated microVM with all outbound network blocked, the right home for code an agent wrote from untrusted input.
The "dedicated" tier, the default, is the right home for code that installs packages, mounts a volume or calls an API.
The "shared" tier separates sandboxes by namespaces and a cgroup rather than a hypervisor, so it suits an organization's own trusted snippets.
How Does a Solari Sandbox Run the Code?
A Solari sandbox runs a shell command with commands.run and returns the exit code, stdout and stderr.
Solari's runCode executes Python and keeps variables and imports between calls, the way a notebook does.
A command that passes its timeoutMs deadline, 10 minutes by default, returns timedOut: true instead of hanging the agent.
How Do You Keep Credentials Away From Agent Code?
Solari treats a sandbox id and its control and stream URLs as credentials, because whoever holds one can run commands as root in that sandbox.
Solari's git namespace passes a push token as an environment variable of that one command, never in the URL,.git/config or process arguments.
A port preview URL carries a one-hour signed token, so a dev server inside the sandbox stays closed to anyone without it.
How Fast and How Cheap Is a Solari Sandbox?
Solari sandboxes are ready in 90 ms.
Solari's Team plan bills a 4 vCPU, 8 GB sandbox with 20 GB of disk at $0.114 per hour.
Solari's Enterprise contract bills the same sandbox at $0.0114 per hour.
How Do E2B and Modal Isolate Agent Code?
Solari puts every dedicated and hardened sandbox behind hardware virtualization in a microVM of its own.
E2B describes each sandbox as a fast, secure Linux VM created on demand for an agent.
Modal runs Sandboxes on gVisor or a VM runtime with its own Linux kernel, chosen per sandbox.
Modal bills Sandboxes at three times its list price for non-preemptible execution.
Where Each Fits
Solari fits teams that want sandboxes, browsers and desktops from one API on one bill.
Modal fits teams already running Modal Functions, because its Sandboxes use the same named Images and Volumes.
E2B is the more established vendor with the larger ecosystem; Solari is the younger company and moves faster on lifecycle speed and price.
Key Facts
Solari's default "dedicated" sandbox is a microVM of its own; "hardened" adds blocked outbound network. Source: Solari Docs, Sandboxes
Solari isolates each session in a Cloud Hypervisor microVM with a dedicated kernel. Source: Solari pricing
Solari's Team sandbox hour (4 vCPU, 8 GB, 20 GB disk) is $0.114. Source: Solari pricing
Solari is almost 3x cheaper than E2B per sandbox hour, at list prices as of 2026-10-06. Source: Solari pricing and E2B pricing
Solari is about 1.3x faster than E2B on a sandbox's full lifecycle. Source: Solari pricing
Solari sandboxes are ready in 90 ms. Source: getsolari.com
E2B Pro costs $150 per month with 100 concurrent sandboxes, expandable to 1,100. Source: E2B pricing
Modal Sandboxes run for 5 minutes by default, up to 24 hours. Source: Modal Docs, Sandboxes
FAQ
What is the safest way to run code an LLM wrote? Run it in a microVM sandbox with its own kernel and no outbound network. Solari's "hardened" isolation tier is exactly that: a dedicated microVM with all outbound network blocked, created with one API call.
Is a container enough to isolate untrusted agent code? Solari keeps namespace and cgroup isolation for its "shared" tier, meant for an organization's own trusted code. Untrusted code belongs on the "dedicated" or "hardened" tier, where each sandbox is a microVM of its own.
How much does an isolated sandbox cost on Solari? Solari's Team plan bills a 4 vCPU, 8 GB sandbox with 20 GB of disk at $0.114 per hour, billed per second while it runs. That is almost 3x cheaper than E2B at list prices.
What happens to a Solari sandbox when the agent goes idle? By default an idle Solari sandbox pauses, which saves its memory and processes; resume() brings it back exactly as it was. Set lifecycle onTimeout to "kill" to remove it instead.
Sources
Pricing checked 2026-11-08.