Skip to content
AI, Security and ISO 27001: How Spadoom Keeps AI Agents in Check
Insights · ·7 min read

AI, Security and ISO 27001: How Spadoom Keeps AI Agents in Check

Dario Pedol

Dario Pedol

Founder & CEO, SAP CX Architect, Spadoom AG

Share

At Spadoom, AI agents do much of the work, and that makes security a daily question rather than an annual one. Our answer has three parts: secrets that agents never see in files or chats, access that is limited and owned by named people, and guardrails that stop rather than bend. Spadoom AG is in the ISO/IEC 27001:2022 certification process with SQS; the Stage 1 audit for initial certification takes place on 24/25 November 2026. This is part 2 of our series on how Spadoom goes AI-driven.

TL;DR: Secrets live in a password manager unlocked with Touch ID, are loaded once per session into memory and never land in files, chats, git or reviewer prompts. Agents get limited access, every system has named maintainers, outgoing messages need a human yes and agents never merge or deploy. When a guardrail blocks something, nobody works around it.

Why AI agents change the security picture

A traditional employee logs in, does a task and logs out. An agent can do hundreds of tasks in an hour, reads large amounts of text written by others and acts on what it reads. That is the point of agents, and it is also the risk.

Three things worry us most:

  • Secrets in the wrong place. An agent that sees a password can repeat it, log it or paste it into the wrong window.
  • Manipulation. Text from an email or a web page can try to give the agent instructions. Part 3 of this series covers prompt injection in detail.
  • Speed. A mistake that a person would make once, an agent can make a hundred times before anyone notices.

Secrets: one place, in memory, never in files

The password manager is the master copy

All passwords, API keys and tokens live in a password manager. It is unlocked with Touch ID. At the start of a session, the variables needed are loaded once into memory. They are not written to files, not pasted into chats and not passed around in commands where they could end up in a log.

No secrets in git

No secrets in any repository, not even in a throwaway script “just for testing”. Configuration files reference where a secret lives, not the secret itself.

Secrets never go to reviewers

Every result at Spadoom is checked by two AI models from two different vendors, as part 6 of this series describes in detail. Secrets are the hard exception: they never go to either reviewer.

At Spadoom, secrets are unlocked with Touch ID, loaded once per session into memory and never written to files, chats or git. Not even our two AI reviewers ever see one.

Convenience is part of security

This sounds strict, and it is. It also has to be comfortable. If security becomes too annoying, people return to plain-text passwords in a notes file. So we designed for one unlock per session, not one prompt per action. Comfort is a security feature.

Access: limited and owned

  • Named maintainers. Every system has named human maintainers. They decide what agents may do in their system and are accountable for it.
  • Technical users. A technical user writes CRM notes from meetings, so it is visible that a system wrote them.
  • No merges, no deploys. Agents open draft pull requests. People merge. Main branches are protected, so changes only reach them through pull requests.
  • No side doors. Agents are blocked from webmail and social networks in the browser, so they cannot bypass the approval for outgoing messages by opening a web page.

Guardrails: stop, do not bend

A guardrail that agents can talk their way around is not a guardrail. Our rule: when a guardrail blocks an action, the agent stops and reports. It does not look for another path.

That rule was tested. Once a guardrail blocked an action we needed. The tempting move would have been to let the agent find a way around it. Instead we did it by hand and kept the guardrail as it was. Rules change deliberately, after a decision, not in the middle of a task.

The four-eyes rule in our content system is another example. In November 2026 it stopped the author from approving his own 136 post revisions. We did not switch it off. We added an audited admin override that requires a written reason, so the exception is visible and can be reviewed.

Lessons from real slips

Most of our security rules for agents did not come from a whitepaper. They came from things that went wrong:

  • Stale credentials. Agents used outdated access data and failed in confusing ways.
  • Rate limits. Agents ran into rate limits of the systems they worked with.
  • Confident claims. Agents reported success that a live check did not confirm. Now we require proof: tests, logs, live checks.

Our ISO 27001 process

Information security is not a side project for us. Spadoom AG is in the ISO/IEC 27001:2022 certification process with SQS; the Stage 1 audit for initial certification takes place on 24/25 November 2026.

For us, agents belong in that management system like any other actor with access: there is a named person responsible, and the rules they work under are part of it.

Frequently asked questions

How does Spadoom keep passwords and keys away from AI agents?

Secrets live in a password manager unlocked with Touch ID. They are loaded once per session into memory and never written to files or chats. There are no secrets in git, and secrets never go to the AI reviewers.

What stops an AI agent from doing something harmful?

Several layers: outgoing messages wait for a human approval, agents are blocked from webmail and social networks in the browser, agents never merge or deploy code, main branches are protected, and every system has named human maintainers.

Where does Spadoom stand with ISO 27001?

Spadoom AG is in the ISO/IEC 27001:2022 certification process with SQS; the Stage 1 audit for initial certification takes place on 24/25 November 2026.

What happens when a guardrail blocks an agent?

The agent stops and reports. It does not work around the block. If the action was legitimate, a person does it by hand or changes the rule deliberately, never in the moment.

Do AI agents fit into an ISO 27001 management system at all?

Yes, if they are treated like any other actor with access: defined permissions, logging, named responsibility and rules that are reviewed. That is how we handle them in our own information security management.

What this means for your SAP project

Before agents get access to your SAP systems, answer the security questions you would ask for a new employee, and a few more. Which technical user does the agent act as, and with which authorisations? Where do its credentials live, and who can see them? Which actions need a human approval? What happens when it is blocked? If your company runs an information security management system, agents belong in it from day one.

We help customers set up SAP Business AI scenarios with these questions answered. See our approach or contact us.

The series: how Spadoom goes AI-driven

  1. AI-driven everywhere: the rules and what we will not automate
  2. AI, security and our ISO 27001 process (this article)
  3. Nothing leaves without a yes: approvals and prompt injection (coming 28 November 2026)
  4. From meeting to CRM without typing (coming 5 December 2026)
  5. We built our own internal systems with AI, and where that is a bad idea (coming 12 December 2026)
  6. Two reviewers, one human: how every result gets checked (coming 19 December 2026)
  7. How AI answer engines see us, and what we changed (coming 26 December 2026)
AISecurityISO 27001Information SecurityGovernance
Ask Spadoom · AI assistant

Ask Spadoom

Answers drawn from what Spadoom has published on this site, with links to the pages they come from.

Try one of these

Enter to send · Shift+Enter for a new line 0 / 600
Continue with an expert Opens the contact form with your question filled in.

AI-generated answers. Verify before acting. Questions are stored anonymously, without your IP address, so we can improve our content. Please do not enter personal data.

Next step

SAP Business AI implementation partner

Spadoom is the SAP Business AI implementation partner across Switzerland, Germany, Austria and Italy. 14-week median go-live. Live customers across DACH.

Related articles

Product Updates 10 min read

What's New in SAP Service Cloud V2 in 2026

The 2026 SAP Service Cloud V2 releases through Q3, sorted by what matters in practice: Agent Inbox, the Digital Service Agent handoff, approvals, item-level complaints, analytics, the Q2 additions from the new email editor to sales objects in the service workspace, and the Q3 Joule AI Units billing change.

David Blumenstock · 2 Jul 2026
Read article →
Product Updates 8 min read

What's New in SAP Sales Cloud V2: Q1 2026 Update

Our field notes on the first 2026 SAP Sales Cloud V2 release: the Joule Studio agent builder reaching general availability, activity-based deal risk signals and reworked forecasting, and what each means for your team.

David Blumenstock · 13 Mar 2026
Read article →