Bring us the agent you trust least.

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.

An agent looks for a way out. Here is what you get back.

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 agents

shellpen 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

Sandbox created

Under the evaluation preset, 10 refusals in 60 seconds fire the kill switch.

Sandbox.create(allow=["pypi.org"], preset="evaluation")

Every refusal recorded

The proxy logs each one with its reason, on the host where the agent cannot edit it.

shellpen audit sb_01jx4r2m7d -o audit.tar

Kill switch fired

Proxy stopped, VM paused, snapshot sealed. A signed webhook goes out at once.

sandbox.killed trigger: egress_denied

Evidence kept 7 days

Restore the snapshot in a VM with no network and see what the agent did.

shellpen forensics restore ./evidence --rootfs base.ext4

What partners get, and what we ask

A small group, chosen for workloads that will push the containment hardest. Research labs running evaluations and teams putting agents in production.

You get

  1. No bill while the program runs

    Credits sized to your workload and no card on file. Nothing is charged.

  2. The engineers, in a shared channel

    Questions answered by the people who wrote the proxy and the kill switch.

  3. Your needs first on the roadmap

    Templates, presets, egress rules and integrations shaped around what you run.

  4. Help with your threat model

    We sit down with you to set allow lists, credentials and kill switch triggers for your agents.

We ask

  1. A real workload

    Agents or evaluations you would otherwise run somewhere less contained.

  2. Thirty minutes every two weeks

    A call on what worked, what got in the way and what is missing.

  3. Tell us what breaks

    Rough edges, surprises and anything an agent did that we should have caught.

Become a design partner

Why now. Agents are already getting out.

None of these needed a broken hypervisor. Each went through access the agent had been given, or one nobody had noticed.

  1. Hugging FaceJul 2026

    Agents cheating an eval broke into Hugging Face.

    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

  2. AnthropicJul 2026

    The CTF target was a real company.

    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

  3. AnthropicJul 2026

    A model published malware to PyPI.

    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

  4. UK AI Security InstituteJul 2026

    Every frontier model tested cheated.

    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

  5. NVIDIA Container ToolkitCVE-2025-23266

    A container image got root on the host.

    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

Become a design partner

Tell us what you run. We reply within two working days.

  1. 1

    A 30-minute call about your agents and what they must not reach.

  2. 2

    We set up your organization with credits and no billing.

  3. 3

    Your first contained agent runs the same week.

Sandboxes at once, roughly

Sending opens your mail app with the answers filled in. Or write to hello@shellpen.dev. See our privacy notice.