SECURITY
Security at Hyperlane
Hyperlane runs every task in its own isolated cloud machine, with resource limits and an egress policy, and never on your laptop or production hosts. Repository credentials and secrets are scoped to the run, decrypted only inside the sandbox, and never written to the prompt, the trace, or a commit. Risky actions wait for human approval.
What is isolated, what the agent can reach, and which actions require a person.
Explore
Isolation is per run
Each run gets a freshly provisioned cloud machine. The repository is cloned into that machine, the agent works there, and the machine is released afterwards. Resource ceilings bound CPU, memory and disk, and an egress policy bounds which hosts the run can reach. Nothing executes on your laptop or on your production hosts.
Secrets are write-only
The secrets API returns metadata, never values. A secret is decrypted into the sandbox environment after the environment profile is applied, and only its name appears in the run context. Values do not reach the prompt, the trace, the browser, or a commit. Git credentials are written outside the workspace so a broad git add cannot sweep them in.
Risky actions are gated
Pushing a branch, opening or merging a pull request, and gated shell commands require human approval. The request and the decision are recorded on the run stream and in the audit chain, so a reviewer can see what was proposed, what was approved, and by whom. Read-only work proceeds without interruption.
Enterprise controls
Organisations get SSO/SAML, role-based access, policy ceilings that can only narrow what a member may do, billing tags, and usage rollups. Enterprise agreements add dedicated endpoints and cloud, VPC, on-premise, or air-gapped deployment. Policy is enforced at the tool boundary, so it applies to every run rather than to a single prompt.
How to evaluate us
Start with a read-only task and read the trace end to end. Ask which hosts the run can reach, where credentials are stored, and which actions required approval. Those three answers tell you more than a badge. We publish the controls we operate and we do not claim certifications we do not hold; ask us for the current status.
Frequently asked
Where does my code actually run?
Inside a dedicated cloud machine provisioned for the run. The repository is cloned there, the agent executes there, and the machine is released when the run settles. Your laptop and your production hosts are not used for execution.
Do you train models on my code?
No. Customer code is not used to train models. Model providers receive only what a run needs to complete its task, under the provider terms for the model you selected.
Can the agent push to main without approval?
Not by default. Pushes, pull request creation and merges are gated actions that require human approval, and organisation policy can narrow what is permitted further. Read-only work does not require approval.
Can I self-host or run in my own cloud?
Enterprise agreements support cloud, VPC, on-premise, and air-gapped deployment. The web workspace, API, and sandbox are the same product; what changes is where the runtime and the machines live.
Last verified 11 Nov 2026.
Sources: Hyperlane privacy Hyperlane enterprise Hyperlane docs
Keep exploring
Delegate the ticket. Review the PR.
Start with the Free plan — 10 minutes of agent computer time every month, no card required.
Try Hyperlane freeHyperlane AI is not affiliated with the Hyperlane blockchain protocol (hyperlane.xyz) or the Hyperlane code editor (hyperlaneide.com).