Security

Every agent and every computer runs in its own Firecracker micro-VM with its own kernel, disk and network. No file system, memory or process is shared between customers. When you delete an agent or a computer, its VM and its disk are destroyed.

The control plane, which is the dashboard, the API and the scheduler, runs on Railway. It holds account, configuration and audit metadata in a Neon Postgres database. Template images and files you upload are stored in Amazon S3. Secrets you give an agent are stored encrypted and passed into its VM.

A person can watch any computer live through a watch link and take control through a takeover link. When an agent reaches a login, a two-factor prompt, a CAPTCHA or a payment, it requests a takeover, and its next action is refused until the person marks the step done.

Every session is logged: when the computer was driven, by which API key, for how long, and how many actions and screenshots it took. Every action returns a screenshot of the result to the caller. To stop an agent, put its computer to sleep or delete it.

Your agent runs with the model and the API keys you choose. Model calls made with your own keys go from the agent to your provider, and Maritime does not store them. If you use Maritime-provided model access instead, calls go through the Maritime proxy to OpenAI or Anthropic. OpenAI processes them under a Business Associate Agreement.

Maritime does not use customer data to train models and does not sell or share it. Both commitments are written into the standard Data Processing Agreement.

Agents and computers run in the United States or in Germany. The German servers are Hetzner servers in Falkenstein. In the EU configuration, agent workloads, their data, logs and backups stay in the European Union.

On the Dedicated plan, your computers run on servers reserved for your account.

Traffic between you and Maritime, including the dashboard, the API and watch and takeover links, is encrypted with TLS. Secrets and API keys you give your agents are encrypted at rest with AES-256-GCM. The platform database on Neon and the object storage on Amazon S3 are encrypted at rest.

Delete an agent or a computer and its VM and disk are destroyed. A sleeping computer keeps its memory snapshot for 7 days after its last action. After that the snapshot is dropped and the next wake rebuilds the computer from its disk.

On written request, Maritime deletes your data within 30 days, removes it from backups within 90 days, and confirms the deletion in writing. The Privacy Policy lists the retention period for each kind of data.

HIPAA. Comp AI issued Maritime's HIPAA attestation on October 19, 2026. Business Associate Agreements are in place with OpenAI, Amazon Web Services, Google Workspace and Neon. Maritime signs a Business Associate Agreement with each healthcare customer.

SOC 2. The SOC 2 program is underway. Policies, the risk register, vendor reviews and evidence are kept in Comp AI.

Agreements. The Data Processing Agreement and the Business Associate Agreement are available on request from contact@maritime.sh.

Maritime gives customers 30 days' written notice before adding or replacing a subprocessor.

Integrations you connect to your agents, such as Slack or Telegram, and model providers you bring your own keys for are your own processors, not Maritime subprocessors.

The Maritime trust centre holds the security policies, the HIPAA certificate and the security questionnaire. Documents are released under NDA.

Email security@maritime.sh with what you found and how to reproduce it. The founders read every report.