Updated 19 November 2026
Trust
Where Bandicot keeps your work, what a company controls and can see, who handles data for us, and what we do and do not certify. Written for a security review.
Where your data goes
Three places can hold something. What each holds is fixed below, and the third holds status only.
On your Macs
- Agents, files, memory, version history and voice samples live on the Mac that made them.
- Prompts go straight from the Mac to the AI provider you picked, under your own agreement with it.
- Messages between teammates are end-to-end encrypted. Our servers hold only sealed copies.
On a cloud computer (optional, Business)
- If a company adds a cloud computer, its automatons run there around the clock, and the work they touch is held on that computer.
- By default the computer is a virtual machine in the company's own Oracle Cloud account, or a Mac or server the company runs. The company is Oracle's customer under its own agreement; Oracle is not a Bandicot subprocessor for it.
- Bandicot holds no Oracle credential and makes no call to Oracle. We learn only what the computer reports about itself, and staff have no access to it.
- A separate mode, off unless we switch it on, hosts a computer on Vercel in the European Union (Frankfurt, Paris, Dublin and Stockholm) or the United States. There any staff access is written down first and shown to the company's admins.
On Bandicot's servers: metadata only
- Your account, your plan and billing, and for companies the people, rules and activity log. Card numbers go to Stripe, never to us.
- For each cloud computer: name, region, size, status, last report time, engine version and how many automatons run.
- For each event that reached a computer: when, how big and whether it arrived. Never the event itself.
- A check in our code reads every cloud table and fails if a column could hold a body, a prompt or a file.
The one exception on a Mac: while Jev is on, short excerpts go to a supplier so it can answer. A company can switch it off for everyone.
Regions and data residency
- A computer in the company's own Oracle Cloud account runs in the home region the company chose when it opened that account. To keep its data in the EU the company picks a European home region; Bandicot neither chooses nor sees it.
- In the optional Vercel mode a computer placed in the European Union stays there: its sandbox, its snapshots and any fallback region are inside the EU (Frankfurt, Paris, Dublin or Stockholm). The United Kingdom is not counted as the EU.
- A company states whether its computers are in the European Union or the United States, and its console and agreement say so. In the Vercel mode every computer of a company is in the same group, and moving groups means removing the computers first.
- Our account database and website are not pinned to the EU today. Account data and the status records above may be processed in the United States, on the safeguards in the data-processing agreement.
Subprocessors
Every company that handles data for Bandicot, as read from what the product actually calls. A business customer is told 30 days before a new one starts and may object.
- Vercel: Hosts bandicot.app and its API and runs the scheduled jobs. Runs cloud computers only in the optional Vercel mode, which is off unless we switch it on.
Handles: Account traffic. In the Vercel mode, the work held on a computer and its snapshots. Where: United States. In the Vercel mode, the region the company picks. - Turso: The database for accounts, plans, company records, the activity log and cloud computer status.
Handles: Account data and metadata, never work content. Where: See the supplier list for the current region. - Stripe: Card payments, invoices and tax.
Handles: Billing details. Card numbers go straight to Stripe. Where: United States and European Union. - Resend: Sign-in links, invites and the weekly digest by email.
Handles: Email address, name and the message. Where: United States. - Apple: Push notifications to iPhone.
Handles: A device token and an encrypted notification. Where: Worldwide. - Google (Firebase Cloud Messaging): Push notifications to Android.
Handles: A device token and an encrypted notification. Where: Worldwide. - Google, GitHub, Microsoft and your own identity provider: Sign-in, only if a person or company chooses it.
Handles: Name and email address at sign-in. Where: Set by the provider. - The Jev supplier: Answers Jev, only while a company leaves it on.
Handles: Short excerpts of what an agent is about to do. Where: Named on the supplier list.
Oracle Cloud is not on this list. A cloud computer in a company's own Oracle account is run by the company under its own agreement with Oracle, so Oracle is the company's processor and not ours. We hold no Oracle credential and make no call to Oracle. The live list, with the notice dates of any new supplier, is on the suppliers page.
Security controls that exist today
Each line names the part of the code that does it.
- Encryption in transit. Every page and API call is served over HTTPS. Team messages are end-to-end encrypted, so the relays that carry them cannot read them. In the code: Vercel TLS, mailbox envelope
- Encrypted secrets. The secret each cloud computer uses to prove itself to us and each event source's signing secret are stored encrypted with AES-256-GCM. A signing secret is shown to the company once. Passwords, tokens and join codes are stored as hashes or sealed, and a join code works once. In the code: cloud-server.ts, bcrypt
- Single sign-on and directory sync. Google Workspace, Microsoft Entra ID and any SAML login, with SCIM adding and removing people from the company directory. In the code: company-sso.ts, company-scim.ts
- Roles. Owner, admin, IT, billing, auditor, group admin and member, each with a short list of what it may do. In the code: company.ts
- Audit log. Every change in the company console is recorded, can be downloaded and can be streamed to the company's security tools. A run log lists each AI run as names and numbers, never the text of anyone's work. In the code: company_activity, company_run
- Retention and legal hold. A company sets how long the logs are kept, or places a legal hold that stops deletion. In the code: setRetention
- Client data shield. Client names, ID numbers and bank numbers are masked before any AI provider sees them. A call that still carries one is blocked, and a check in Versions turns red. In the code: company-shield
- Webhook signatures. An event reaches a cloud computer only if it carries a valid HMAC signature made from that source's own secret. A wrong company, computer, source or signature all get the same answer. In the code: cloud-server.ts
- Break-glass access log. Bandicot staff cannot reach a computer in the company's own Oracle account or on its own machine: we hold no credential for it. For a computer we host in the optional Vercel mode, staff reach it only after writing the reason, and the entry appears in the company's console at once. In the code: cloud_operator_access
- Passkey-protected administration. Our own admin area needs a passkey, and every administrative action is recorded. In the code: admin-auth.ts
- Deletion. Deleting an account removes it and cancels any paid plan. Removing a cloud computer forgets its record here at once; the virtual machine in a company's own Oracle account stays the company's to destroy, and a computer we host is deleted with its snapshots. A lapsed plan's computers are stopped within minutes and their records or hosted copies are deleted 60 days after its last paid day. In the code: keepAlive, purgeLapsed
Privacy law and AI
- GDPR and UK GDPR. For a company's people, the company is the controller and Bandicot its processor. The data-processing agreement is accepted online, names our security measures, a 48-hour incident notice, supplier objection rights, deletion at the end and the transfer safeguards (the EU-US Data Privacy Framework where a supplier is certified, otherwise the Standard Contractual Clauses).
- CCPA and other US state laws. For a company's data we act as a service provider or processor. We do not sell personal information, do not share it for advertising and never use it to train AI models. Requests to see, correct or delete data go to the contact below.
- EU AI Act. Bandicot is a tool that runs AI models chosen by the user, and everything an agent writes, says or draws is made by AI. Images carry a machine-readable mark in the file. Bandicot is not built for hiring, credit, schooling or other high-risk uses, and a company using it there makes its own assessment first.
Certifications
We do not hold a security certificate today, and we say so.
- SOC 2 Type II: In preparation. No report has been issued.
- ISO/IEC 27001: Planned. Not certified.
Each supplier above publishes its own certifications. When one of ours is issued, it will appear here with its date and the auditor's name.
Paperwork
Report a problem
Security reports, privacy requests and questions from a security review go to the address below. We answer within two working days.
Bandicot is made by OG BUSINESS LLC.
Also: Trust · Terms · Privacy · Cookies · Support · Suppliers · Data processing