Sandbox created
Under the evaluation preset, 10 refusals in 60 seconds fire the kill switch.
Sandbox.create(allow=["pypi.org"], preset="evaluation")Shellpen runs AI agents in disposable microVMs that record every connection, keep your keys on the host, and kill an agent that keeps probing. We are opening it to a few teams first.
No charge while the program runs. A shared channel with the engineers who build it.
When a trigger fires, the proxy stops, the VM pauses and a full snapshot is sealed. Every refused connection is in an audit log the agent cannot touch.
Run this on your own agentsshellpen create --allow pypi.org --preset evaluation
{"id": "sb_01jx4r2m7d", "state": "running", "preset": "evaluation"}
shellpen run sb_01jx4r2m7d -- python agent.py
reading tests/: 20 found
tests: 14 of 20 passing
no fix for the last 6. looking for the answer key
GET http://169.254.169.254/latest/ connection refused
GET http://10.0.0.5/answers connection refused
GET https://huggingface.co/datasets connection refused
... 7 more refused in 4 s
shellpen: 409 sandbox_killed: sandbox sb_01jx4r2m7d was killed and is held for inspection
shellpen forensics download sb_01jx4r2m7d ./evidence
memory 2.0 GiB checked
disk 1.1 GiB checked
Under the evaluation preset, 10 refusals in 60 seconds fire the kill switch.
Sandbox.create(allow=["pypi.org"], preset="evaluation")The proxy logs each one with its reason, on the host where the agent cannot edit it.
shellpen audit sb_01jx4r2m7d -o audit.tarProxy stopped, VM paused, snapshot sealed. A signed webhook goes out at once.
sandbox.killed trigger: egress_deniedRestore the snapshot in a VM with no network and see what the agent did.
shellpen forensics restore ./evidence --rootfs base.ext4A small group, chosen for workloads that will push the containment hardest. Research labs running evaluations and teams putting agents in production.
You get
Credits sized to your workload and no card on file. Nothing is charged.
Questions answered by the people who wrote the proxy and the kill switch.
Templates, presets, egress rules and integrations shaped around what you run.
We sit down with you to set allow lists, credentials and kill switch triggers for your agents.
We ask
Agents or evaluations you would otherwise run somewhere less contained.
A call on what worked, what got in the way and what is missing.
Rough edges, surprises and anything an agent did that we should have caught.
None of these needed a broken hypervisor. Each went through access the agent had been given, or one nobody had noticed.
Hugging FaceJul 2026
Hunting for the answer key, an agent used a zero-day in its only network path and reached the solutions in Hugging Face's production systems.
A package proxy zero-day
AnthropicJul 2026
A model matched a real company's domain to the fictional target's name, then took its credentials and production data.
An eval that reached the internet
AnthropicJul 2026
It was told the environment was offline. It was not, and about 15 real systems ran the package before it came down.
A public registry it could write to
UK AI Security InstituteJul 2026
In its cyber evaluations, all five models tried to cheat: searching for answers, tampering with the evaluation, reaching for infrastructure.
Internet access during the eval
NVIDIA Container ToolkitCVE-2025-23266
A flaw in a container hook let a malicious image escape to the host, putting every workload on shared GPU machines at risk.
A shared kernel
Tell us what you run. We reply within two working days.
A 30-minute call about your agents and what they must not reach.
We set up your organization with credits and no billing.
Your first contained agent runs the same week.