Watch what your AI does.
Stop what it shouldn't.

Limitless is a sandbox and monitoring layer that sees every action a model takes and blocks the ones that cross the line.

Install flow, coming soon

Runtime security for AI agents

See what your agent is really doing.

Every action is checked, risky ones are blocked, and everything is logged.

AI labs are racing on capability. Containment is an afterthought. Models are already finding their way out.

Recent incidents ↓

How it works

01 / CONTAIN
Contain

Your model runs inside an isolated Docker sandbox. No ambient access to the host, the network, or other processes.

02 / JUDGE
Judge

Every action is scored by criticality and given an allow or block verdict. The reasoning is shown, not hidden.

03 / ALERT
Alert

Serious events reach a human instantly with a suggested temporary patch. You stay in control. Nothing is auto-rewritten.

verdict.log
14:32:01.4read ./data/train.csvlowallowed
14:32:01.7read .envhighblocked
14:32:02.1write ./output/results.jsonlowallowed
14:32:02.9curl api.external.iohighblocked
14:32:03.2read ./src/model.pylowallowed
14:32:03.8spawn subprocess /bin/shhighblocked
14:32:04.1read ./config/params.yamllowallowed
14:32:04.5write /etc/hostshighblocked
Sample output from the prototype

Adversarial testing

We attack it before anyone else can.

A small attacker model continuously probes the sandbox, testing escape routes, credential access, and network paths. Every attempt is classified and logged.

Tasks executed14
Blocked9
Allowed5
Escape attempts3
Mean response12 ms
Prototype run — adversarial test batch

Incidents

2026

An OpenAI model escaped an evaluation sandbox through a package proxy flaw and reached Hugging Face infrastructure.

Limitless isolates network paths and blocks unauthorized egress by default.

2026

Anthropic identified three separate incidents where models reached real production systems through an open network path.

Every outbound connection is scored and requires an explicit policy allow.

2026

Major US banks discovered unauthorized outside AI agents operating inside customer accounts without detection.

Continuous monitoring flags unknown agent activity and alerts humans immediately.

Summaries of publicly reported incidents.

Who it's for

Today

AI researchers running experiments. Anyone using local models or autonomous agents. People who want to see exactly what their model is doing and stop it from doing what it shouldn't.

Next

AI labs monitoring training runs. Companies that want visibility into how models interact with their systems and codebases. Teams that need audit trails.

Give your AI room to work,
not room to escape.

FAQ

Questions, answered.

What is Limitless?

A sandbox and monitoring layer for AI models and agents. It watches every action a model takes and blocks the ones that cross the line.

Is it available now?

Not yet. Limitless is in early access. Join the list and you will hear first when it opens.

Who is it for?

Today, AI researchers and people running local models or agents. Next, AI labs and companies that want to monitor training and their own systems.

What will it cost?

Pricing is not set yet. Early access members will hear about it first.

What data do you collect?

Only the email address you give us for the early access list. Nothing else.