Security
How the platform is built to fail safely: isolation per agent, credentials that never reach a model, an approval gate the model cannot move, and a log of everything.
Zuger Sealed
Your workspace is Sealed.
Sealed by default. Opened only with your approval.
Seal Log
- Run title
- Run steps
- Resealed
Seal request1 h
Ruth at Zuger asks to read one Run
Ticket 4411: the August close stops at the bank step.
Sealed. Zuger’s staff tools show counts, dates and status, never your work.A named person at Zuger asks to read one Run, for an hour, and says why.You approve with your passkey. Each look is listed as it happens.Time is up. Resealed, on its own, and you are told.
Closed by default
Zuger’s staff tools show counts, dates and status. Your documents, Runs and answers stay redacted.
Opened only with your approval
A named person asks, with a reason and a scope. An owner or admin approves with a passkey, for 15 minutes, 1 hour or 4 hours, or denies it.
Every look, listed
Each value opened is written to your Seal Log before it is shown. Access ends on time, or the moment you reseal.
Sealed covers access by Zuger’s staff. Your documents are still processed by Zuger agents to do the work you ask for.
Zuger Sealed
Zuger’s staff tools show counts, dates and status, never your content. To read anything, a named person at Zuger has to ask, and an owner or admin at your company has to say yes.
- They ask, you decide. Each request names the person, the reason, and exactly what: one Run or named documents, for 15 minutes, 1 hour or 4 hours. You approve it with your passkey, shorten it, or deny it. A request nobody answers lapses after 72 hours.
- Read-only, then closed. Approved content opens in a watermarked, read-only viewer with copy and print switched off. It closes when the time runs out, or within seconds of you ending it.
- Every look, listed. Every value opened is listed in your Seal Log in the Console, by name and time, and you can export it. You are emailed when someone asks, when access opens and when it ends.
- No machine can say yes. An API key can never approve a request or read content.
- Our logs carry no content. Logs and error reports keep a fingerprint of a value, never the value.
- Encrypted with your company’s own key. Your documents, Runs and answers are stored encrypted under a key that belongs to your company alone. Those keys are protected by a master key kept in our hosting provider’s secret store, not in a separate key service. That protects a copy of the database or a backup. It does not stop someone with access to the running servers or to that store: the access controls below are what cover them.
What Sealed does not cover. Agents read your content to do the work you brief them on, and our model provider processes it under terms that do not allow training on it — see Subprocessors.
The engineers who run the infrastructure are covered by the access controls below, not by Sealed.
The threat we design against
The interesting failure is not a leaked database. It is an agent doing something plausible and wrong with real consequences — because a web page told it to, or because it inferred a permission it did not have. Two structural choices make that outcome boring: agents prepare but never file, sign or pay, and no agent holds a credential that can move money. Today nothing in Zuger can move money: Agent money is built and switched off, and when it opens, money moves only after a person approves it with a passkey, never on an agent’s say-so.
Isolation
Isolation is at the data and credential layer, not the machine layer: an agent never holds a token. It reads what your company has synced, through a fixed list of read-only look-ups for its job, and the query layer, not a prompt, keeps it inside your company. There are no per-agent virtual machines and no browser automation — see the limits below.
Credentials
Scoped OAuth tokens, encrypted at rest with per-tenant keys, decrypted only inside a separate connector service, which agents never reach. They never enter a prompt, never appear in a transcript, never reach a subprocessor and are never returned to a browser. Disconnect a tool and its token is deleted at once, and nothing new comes in from it. What was already synced stays with your company until you delete it. To withdraw the grant at the provider as well, remove Zuger in that tool’s own settings. We never hold your bank credentials: today you upload a statement, so Zuger never touches your bank. When live feeds open they will be read-only, through a licensed open-banking provider.
Connections coming online
The connections coming online, such as Qonto, Mercury, Deel and Intercom, are not live yet. All of them read only, except Stripe Treasury and Link’s agent payments, coming soon with Agent money, which will pay only what you approved. The Trust page lists each one.
- When these connections open, account and card numbers will be masked to the last four digits before any agent reads them.
- Salary data from them will be shown only to people whose role already allows it.
- As the new connections come online, each will be checked daily. If a provider adds a new action, Zuger will block it until we’ve reviewed it.
Zuger can’t move money, run payroll, sign or file through any connection.
Prompt injection
Content an agent reads — a web page, an email, a PDF, a form — is data, never instruction. Mandates and thresholds change only through an authenticated human action outside the run. An instruction found in content that tries to change either has no effect, and Bernina is designed to log it and show it to you as an attempt.
The approval gate
Enforced in the platform, not in the model. Money out, filings, contracts and outbound communication stop and wait, with no exceptions today. Letting Moléson answer simple questions on its own is coming. An agent cannot change what stops it, approve its own request, or ask another agent to approve it. Approval requires a human session token; a machine token can read the queue and nothing more.
Pilatus doesn’t connect to Stripe: it prepares the Stripe Atlas application for you to enter and submit yourself.
Audit
Every Run keeps its own trail: each tool call, each document read and each output it prepared, step by step, under the run id. Every approval and decision, with who made it and when, every message drafted or sent in your name, every change to a connection and every look by Zuger staff goes into the company’s audit log, which is append-only and hash-chained, so an edit after the fact shows. Zuger Max and Ultra export it in full. Zuger’s staff tools show content only through a Zuger Sealed request you approved, and every value opened is listed in your Seal Log. In the database itself, your content is stored sealed.
Emergency stop
Coming soon. One action stops an agent, the agents that hold one skill, or your company’s whole fleet.
- Who can stop. An owner or an admin, signed in, with a passkey made for that one stop. Never an API key or a connected app.
- Two people for the fleet, if you want. The owner can ask that a stop of the whole fleet waits up to 15 minutes for a second owner’s or admin’s passkey. When nobody else could give one, it goes through alone, and the record says so.
- What it does. The agents’ credentials and keys are revoked, their sessions end, Runs under way are cancelled and never charged, and open proposals are withdrawn. The addresses you subscribed are told at once.
- Who sees it. Sites and apps that check Zuger credentials see the stop within the status list’s window. Anything that does not check cannot be reached.
- Only a passkey restarts. The agents come back as they were, and the restart is recorded.
- Time to stop: Not yet measured. It will be measured by a monthly drill, and shown only once measured.
Access control
Least privilege and named approval. Zuger OPS, our staff tool, needs a passkey for every sign-in and every action, and every staff action is recorded. Production infrastructure is reached only by the founder.
Testing
Internal security reviews of sign-in, tenant isolation, billing, the agents and Zuger OPS were completed in September and October 2026. A third-party penetration test is planned; its summary will be available under NDA. Every change runs the full test suite and a scan for leaked secrets before it can ship.
Resilience
Database backups every day and every week, plus our own encrypted copy every hour, with a written restore procedure and a scripted drill that checks a restored copy. Runs are resumable: an interrupted run reports or resumes, and never half-commits an action against a registry.
What we do not build
No per-agent virtual machines, no browser automation, no stored third-party logins. If a capability would require driving a website with your password, it is out of scope — permanently, not until version two.
Reporting a vulnerability
Write to security@zuger.ai. Until that mailbox can receive mail, use the contact form and choose “Security review”. Acknowledged within one working day, triaged within three, and you will hear the outcome. We do not take legal action against good-faith research, and we will credit you if you want us to.
Questions about this?
Write to privacy@zuger.ai for anything on this page. A person answers, not a portal.