Trust

Security at Rykon AI

An agent that can act needs limits it cannot talk its way past. This is how Rykon AI keeps them, written from what the service actually does, with nothing claimed that the code does not prove.

Last reviewed
1 November 2026
Data location
By provider, on the subprocessors page
Report a vulnerability
support@rykon.ai

A fresh sandbox for anything that runs code

Simple answers come straight from the control plane, which runs no code the model provides. Anything that runs code, a shell, files or packages gets exactly what one task needs, and nothing that would let it reach further. The limits are in the system, not in a prompt.

Two lanes

A run answered directly works in the control plane with no tool that runs code. Work that needs code, a shell, files or packages becomes a child run in a sandbox, asked for with a reason, checked against the agent’s tools and kept on the record.

Fresh, then gone

Sandbox work starts in a new sandbox made for it, kept only for the same task’s next sandbox step, and removed after it. A sweep removes any sandbox that is no longer needed, and fails runs past their deadline.

Locked down

A non-root user, a read-only root filesystem with scratch space only, no Linux capabilities, no privilege escalation, the default seccomp profile, and fixed memory, processor and process limits. On Kubernetes, a gVisor runtime and no service account token.

One address it can reach

A sandbox’s network reaches the relay and nothing else: not the internet, not cloud metadata, not our databases, not other runs. The runtime starts no work until it has checked that forbidden addresses do not answer.

Strict isolation, when you want it

An owner or admin can set the account, or a single role, to Strict: every run, plain replies included, then executes in a sandbox, at the cost of a few seconds and more compute per reply. It is available on paid plans and can be part of an enterprise contract, which keeps it on.

Tools run outside

An agent’s tools never run in its sandbox. The agent asks; the control plane decides by the agent’s level and runs what is allowed.

Keys and credentials

What we do not keep cannot leak. What we must keep is sealed and used only where it is needed.

One credential per sandbox run

Sandbox work holds a single token that works only while it runs, before its deadline, for an active account, and only from its own sandbox address. Copied anywhere else, it is refused. It is stored only as a hash. A run answered directly holds none: it never leaves the control plane.

Your model key never enters a run

Each account’s AI gateway key is sealed with AES-256-GCM, bound to the account, and opened for one model call at a time, by the relay for sandbox work or by the control plane for a direct answer, on the run’s behalf. It is never placed in a sandbox.

API keys shown once

A key is shown exactly once and stored only as a SHA-256 hash. It holds only the scopes it was given, never more than the person who made it, and is read on every request, so a revocation applies to the very next one.

Nothing secret in transcripts

Anything shaped like a key or token is redacted from a run’s transcript before it is stored, and the record keeps a model call’s tokens and cost, never its content.

People decide

Agents act inside a level you set. Above it, the exact action waits for a person.

Risk classes and levels

Every tool carries a class from R0 (read, report) to R4 (outbound to strangers, money, deletion, commitments). An agent acts alone only up to its level’s ceiling, and starts at L0, where everything but reading and reporting waits.

R4 always waits

Whatever an agent’s level, an action in the highest class waits for a person. Rules you add can only ask for more approval, never less.

The exact action, editable

An approval shows the action as it will run. A person may edit it; what the agent proposed is kept beside it, and the edited action is the one that runs, once.

Only a person

An approval is decided by a person signed in to the console. An API key is refused, whatever its scopes.

Accounts and access

Every request is scoped to one account, decided on our servers, on every request.

Walled-off accounts

An account is resolved from the key or person on every request, never from input. A correct id from another account reads as one that does not exist, and the attempt is recorded.

Roles

Owners, admins, members and viewers each hold a fixed set of permissions. Raising an agent’s level and releasing a kill switch are for owners and admins only.

Our own sign-in

Sign-in runs on an identity service we host, with passkeys and authenticator apps, email confirmation before an account works, and a check that turns away a source after repeated failures. The website receives a signed token using PKCE, state and nonce checks, and never sees your password.

No keys in the browser

The console calls the control plane from the server only. Every change is a named server route that checks the session, then refuses a request from any other origin. A build check fails if anything shaped like a key reaches the code a browser downloads.

Stops and ceilings

Work can be stopped at any scope, and spend has a ceiling enforced before the call.

Kill switches

Stop one run, one agent or the whole account. While a switch is on, every model and tool call it covers is refused, approved actions do not run, and live runs are stopped in the next sweep, which engaging starts at once.

Daily spend ceiling

Model calls go through our AI gateway on a key with its own daily ceiling, checked before each call. If the limiter cannot answer, paid calls are refused rather than let through unmetered.

Deadlines and limits

Every run has a deadline and a turn limit. A run past its deadline is stopped and failed, and any sandbox it used is removed.

Records and recovery

What happened is written down where it cannot be rewritten, and the data can be brought back.

Two append-only logs

The record (what agents did and what was decided) and the audit log (who did what to the account, including every refusal). The database refuses to change or delete an entry in either, and each is a hash chain. The one exception is erasing a person or a closed account: names, addresses and content in their entries are replaced under a recorded erasure, and the chain still proves that nothing else changed.

Replicated database

The database cluster runs on two servers in two separate data centres, one a synchronous standby of the other.

Encrypted offsite backups, tested

Backups are encrypted before they leave our servers and stored offsite. The whole database cluster is restored from them once a month to prove they work.

Where your data lives, named

Our own servers and every provider we use are listed with where they run on our subprocessors page, and a new region is listed there before it holds customer data.

What we do not claim

Rykon AI does not hold SOC 2, ISO 27001 or any other certification today, and has not been through an external audit or penetration test. We will say so here when that changes, and not before. Our data processing addendum sets out how customers can review our measures in the meantime.

Reporting a vulnerability

If you believe you have found a security issue, write to support@rykon.ai with “Security” in the subject, and give us a reasonable time to fix it before telling anyone else. Please do not access other people’s data, disrupt the service, or run work that is not yours as part of your testing. We will not take action against research done in good faith under these rules.

Security · Rykon AI