Compliance documentISO 27001 A.5.1

Information Security Policy

What Stilops wants to achieve with information security, who decides what, and which rules apply — including what the policy has to admit.

This is an automatic translation of the Dutch document. In case of any difference, the Dutch text prevails. Read the Dutch text

Version12-11-2026LanguageEnglishPDFText file

Control iso27001:A.5.1 in the compliance register; also supports clause 5 (leadership), iso27001:A.6.7 (remote working) and iso27001:A.8.24 (cryptography).

Adopted by the owner on 2-10-2026. At the same time, the owner accepted the residual risks that the risk register treats as "accepted", including one person acting as owner and administrator and the absence of an external penetration test.

This is the policy that ISO 27001 A.5.1 requires: a single document stating what Stilops wants to achieve with information security, who decides what, and which rules apply. It is deliberately short and it describes what actually happens. A policy that is nicer than the practice is not a policy but a wish.


1. Purpose and scope

Stilops provides AI agents that perform work on behalf of an organisation. Customers entrust us with their business data and sometimes with personal data of their own relations. The core of the product is not only that the work gets done, but that it can afterwards be traced what happened and who approved it.

This policy applies to the entire platform: the customer portal, the administration environment, the public site, the execution of agent tasks, and the infrastructure on which all of that runs. It applies to everyone who works on it — at present the owner — and to the agents that themselves work in the codebase.

2. Objectives

  1. Confidentiality. The data of one organisation is not accessible to another organisation, and secrets of connected systems are stored encrypted and are not shown to an agent. Where an agent technically needs a secret, it receives it only for the duration of that task; the exceptions are set out under "Cryptography and key management" and in §5.
  2. Integrity and traceability. What an agent proposed, who approved it and what the result was, is recorded and is not changed afterwards.
  3. Availability. The service is running, and a disruption can be recovered from without loss of the audit trail.
  4. Human authority. No effect without a human having approved it.

The fourth is at Stilops not a security measure but the starting point of the product. The other three exist to make that fourth one credible.

3. Who decides what

SubjectWho
Adopting and amending the policyThe owner
Security assessment, threat intelligence, incident analysisThe CISO role
Anything that touches a legal obligation or becomes a public statementThe owner, always — this cannot be delegated and cannot be switched off by mandate
Technical implementation, changes to the softwareThe development roles, through the fixed quality gate
Granting authorisations to an agentThe owner

That the owner is the only human in this table is a fact and not a shortcoming to be hidden — but it is the single largest risk in the risk register, and it is recorded there as such.

4. Rules that apply

Access. Access is by name and by role. Passwords are stored as a strong, salted hash. Two-factor authentication is available to every user, and is mandatory as an additional step for actions with consequences — including publicly confirming a compliance claim. On the administration side, a valid two-factor code opens a window of 15 minutes on that session, in which subsequent high-risk decisions do not require a new code. Irreversible actions, such as disabling the two-factor authentication of a user, always require a fresh code, without a window. There is a rate limit on logging in and on forms.

Workstation and remote working. Stilops has no office; all administration takes place remotely, from the owner's workstation. The following rules apply: the disk of the device is encrypted and the screen locks automatically after a short period of inactivity; the operating system and browser receive security updates as soon as they are released; the accounts with the code, hosting, DNS and payment providers are secured with two-factor authentication; access to servers is only via SSH keys, not via passwords; production data is not copied to the workstation, except a temporary copy for a recovery or an investigation, which is deleted afterwards. In the event of loss or theft of the device, the sessions are terminated within 24 hours and the keys that were on it are renewed, and it is assessed whether it is a personal data breach under the Incident and data breach procedure.

Secrets. Secrets of connected systems and of the infrastructure are stored encrypted (AES-256-GCM) with a key per environment, and are decrypted only at the moment a task needs them. They never end up in the prompt to the language model. Accidentally committing a secret is only partly prevented: a check in the quality gate searches the code for secrets with a fixed prefix (such as GitHub, Anthropic and Stripe keys and private keys), and the same patterns are applied to what agents submit as a change. A secret without a recognisable prefix, such as a Hetzner token or a password, is not caught by that check. A user who puts a secret in a chat message has that message intercepted before the model sees it.

Execution of agent tasks. An agent task runs contained under its own, restricted account from a fixed pool. If there are more concurrent tasks than the size of the pool, a task waits for a free account, by default for at most a quarter of an hour, and then does not start. There is no fallback to a shared account or a shared directory, so that a task can never read the decrypted configuration of another task. An agent of a customer never gets access to data of another organisation. It has no own mailbox at Stilops. If the customer itself connects its e-mail, calendar or files (for example Google Workspace or Microsoft 365), the agent can read those. It can send mail or change anything else only if the customer has permitted writing for that connection, and then only in a task that a user of that customer has approved in advance.

Changes to the software. Every change goes through a fixed automated gate: style, type checking, the complete test suite, a check for committed secrets and a check for known vulnerabilities in dependencies. A change to a gate or an authorisation must moreover be deliberately broken to prove that the corresponding test turns red — a gate that has not been broken guards nothing until the contrary has been demonstrated.

Changes and compliance. Every change is also assessed, before merging, for what it means for compliance: does it touch anything that a compliance document, the controls or the terms promise? Software determines which subjects a change touches; a separate, read-only review agent then assesses whether the change may go through as it is, whether a compliance document must change along (which Stilops itself updates in the same change, after which the owner adopts the new text), whether the owner must decide, or whether the change is blocked. A change to the terms, the privacy statement, the data processing agreement or the list of sub-processors, a risk that must be accepted and a measure that becomes weaker, always lie with the owner; so does a check that does not reach a verdict. In addition, every quarter Stilops lays all compliance documents alongside the code and turns every deviation into a finding.

Logging. Actions of administrators and of organisation users are recorded append-only. A recorded action is not changed and not deleted.

Cryptography and key management. Traffic runs over TLS. Data at rest that constitutes a secret is encrypted with AES-256-GCM, with three known exceptions: the short-lived read token of the Search Console connection (valid for at most an hour) is deliberately stored unencrypted; the Anthropic API key for customer runs is present as a file on disk during such a run, until the agent gives its first output or terminates; a leftover file is cleaned up by the next start after two minutes, but a task that hangs without executing anything keeps the file for as long as it hangs — there is no hard upper limit; and the two internal shared secrets for draining on a rollout and for landing a worktree are stored as plain text in the database as long as they have not been set in the environment of the service. That database is readable on production for the agent account, and with those secrets one can abort running tasks and push changes. That has to go: it is not an accepted residual risk but an open deviation. The key is stored per environment outside the source code and the database and is readable only by the service itself, on machines where the agents run under a separate account (a separate setting for that account). Where that separation has not been set up, such as on the local development and production environment, agents run as the same user as the service and can read both the key and the encrypted values.

What an agent can read, even where the separation is set up: the Anthropic API key during a customer run; the ordinary Cloudflare token; as long as they have not been set in the environment of the service, the two internal shared secrets for draining on a rollout and for landing a worktree; as long as the local tunnel for secrets of connections is not switched on, the decrypted configuration file of its own task. Where that tunnel is switched on, those secrets do not end up in a readable file, with one exception: a task of a customer with a connection (the connection bridge, the route tool or a customer's own MCP server) does not yet run via the tunnel. Its configuration file with the decrypted connection secrets is in the own directory of the account under which that task runs. Another task cannot read it. The customer's own agent is kept from reading it by the fixed check on its reading, not by the operating system. That is not an accepted residual risk but an open deviation that is being closed. The database with the data of all organisations is also readable for the agent account, because the servers that agents use need it. An agent of a customer cannot get to it: it then has no shell, a fixed check before every action restricts its reading to its own working directory and attachments, and the tools it receives filter on its organisation. That separation is in the application, not in the operating system: if that check falls away, nothing stops the account. Apart from the two internal shared secrets and the connection secrets of customer tasks, these are residual risks, and the first is recorded in the risk register (R13); those are an open deviation.

External access keys are renewed to the extent the supplier allows. The Cloudflare token is renewed by Stilops itself, only on the production server and only if a separate rotation token has been set that may exclusively manage tokens. The ordinary token deliberately does not get that right, because agents can read it. If the rotation token is missing or Cloudflare refuses, the old token remains valid, Stilops creates a finding and tries again only after a full interval. Until then the token is not renewed. For Hetzner, GitHub, Anthropic, Stripe and MXroute, Stilops creates a reminder as a finding and the owner renews the key himself. For the other connections (S3, OneDrive, VirusTotal, LinkedIn, PageSpeed, Bing, the Google service account and the SMTP and IMAP passwords) no reminder is in place; there only the manual key rotation in the operational planning applies. For Hetzner there is in addition a monitor that reports every action in the action log of Hetzner that does not originate from Stilops itself; for the other suppliers that monitoring does not exist.

A key change after a (suspected) leak invalidates existing public compliance confirmations: those must then be given again, because whoever had the old key may have forged them. In the case of an ordinary change, Stilops re-signs every confirmation that was valid with the old key with the new key. Sessions and invitation and reset links expire on every change.

Incidents. What an incident is, how it comes in and within what period what is reported to whom is set out in the Incident and data breach procedure. The decision to report is human.

Suppliers. Every party that processes personal data is listed by name, purpose, country and the status of the processing agreement in the Sub-processor list — also where that agreement is missing.

5. What this policy must honestly admit

  • There is no segregation of duties. One person adopts policy, approves and administers. The compensation is that the audit trail is complete and append-only, and that the heaviest actions require an extra verification step. That is a compensation, not a solution.
  • Recovery is tested, but not yet completely. Decrypting and checking the back-ups has been tested; a full recovery exercise per quarter has been set up and the first execution is still open. Details are in the continuity plan, which customers can view in the portal.
  • Secrets are not unreadable to every agent. See "Cryptography and key management": the Anthropic API key during a customer run, the ordinary Cloudflare token, and on machines without a separate agent account also the key itself. Sealing those off is an open trade-off. In addition, the two internal shared secrets for draining on a rollout and for landing a worktree are stored as plain text in the database as long as they have not been set in the environment of the service, and that database is readable on production for the agent account. That is not a trade-off but an open deviation that has to go.
  • The separation between organisations in the database is a separation in the application. The database with the data of all organisations is readable for the agent account, because the servers that agents use need it. That an agent of a customer cannot get to it is arranged by a fixed check in the application, not by the operating system: if that check falls away, nothing stops the account.
  • There is no external penetration test.

These points are recorded as an open point in the compliance register. They are stated here because a security policy that names only the strong points misleads the reader about what he can rely on.

6. Compliance and review

The compliance register at stilops.com/nalevingsregister is the ongoing measurement of this policy: per standard and per control what the status is, with evidence. Deviations come in as a finding and are routed to whoever is responsible for them.

This policy is reviewed when the organisation grows to more than one person, when there is a new category of data or a new supplier, after an incident with impact, and at the latest annually.

Questions about this document? Email [email protected].