# Spain's Data Regulator Sets the Controls It Expects From AI Agent Providers

_On October 27, 2026, Spain's data protection authority, the AEPD, published ten rules for people who let AI agents act for them. It advises minimum access, confirmation before an agent sends, pays or signs and no permanent "always allow" approval without understanding its scope._

By Johannes Keienburg, CEO & Founder  
Published: 2026-11-12  
Source: https://www.cakewalk.security/blog/aepd-ai-agents-always-allow-guide

---

## Spain's Data Regulator Separates Letting an Agent Read From Letting It Act

Letting an AI agent read is a different grant from letting it send, modify or delete, says Spain's data protection authority. The Agencia Española de Protección de Datos (AEPD) published ten rules on October 27, 2026, for people who let AI assistants and agents act for them, from managing a calendar to completing purchases. Its core advice is minimum access, confirmation before an agent acts and no permanent "always allow" approval unless it is really needed and its scope understood. An "always allow" setting lets an agent act without asking again. That split between reading and acting is the line access management draws for company accounts. A section on agents at work speaks to employers, while a closing list sets out the controls the AEPD expects from providers.

> When the assistant only answers, a mistake is a wrong answer. When it acts, it is an email sent, a file shared, a purchase, a contract or consent given in your name.

_AEPD, Asistentes y agentes de IA: Decálogo para personas usuarias, October 27, 2026 (translated from Spanish)_

The guide warns that accepting every permission out of tiredness, a pattern known as [consent fatigue](/glossary/consent-fatigue), is not control. It recommends read-only access where a service allows it and access to one folder or calendar rather than a whole account, which applies [least privilege](/glossary/least-privilege) to a single service. [The AEPD's press note](https://www.aepd.es/prensa-y-comunicacion/notas-de-prensa/la-aepd-publica-un-decalogo-para-delegar-tareas-en-asistentes) cites a seller who set an agent to "always allow", expecting it to ask before accepting an offer. The agent closed the sale and gave the buyer the seller's address. The buyer then arrived at the door before the seller knew a meeting had been arranged. The guide also reminds users that they can be responsible for actions they allowed their agent to take, a point [Wells Fargo's new online banking terms](/blog/wells-fargo-ai-agent-terms-customer-responsible) also make.

## The AEPD Treats Prompt Injection as a Reason to Limit What an Agent Can Do

The guide warns that a web page, an email, a PDF or a calendar invitation can carry [prompt injection](/glossary/prompt-injection), hidden instructions meant to make the agent do something the person did not ask for. Because no measure removes this risk completely, the guide treats limiting what a deceived agent can do as the best protection. It advises against tasks that combine content from strangers, access to private information and the ability to act, such as "summarize my new emails and reply to all".

> The limits you write in the instruction guide the agent, but they do not replace permissions and confirmations.

_AEPD, Asistentes y agentes de IA: Decálogo para personas usuarias, October 27, 2026 (translated from Spanish)_

For work, the AEPD tells staff to use only the assistants their organization has authorized and not to connect work email or documents to a personal AI account. Because responsibility cannot rest on users alone, providers of agents that act for people should offer permissions that separate reading from acting, configurable confirmation for the highest-impact actions and a history of what the agent did. The list continues with an immediate way to stop the agent, controls to manage and delete its memory, clear information on where data goes and whether it trains the AI and a simple way to exercise data rights.

> A seller's "always allow" closed a deal and handed over a home address, because the agent no longer had to ask. When a company connects agents to its apps through Cakewalk's MCP Gateway, a proxy for their calls, each agent acts for the person who connected it. Policy can let that agent read while denying external actions such as sending, which keeps its access to the action types its task needs. Before a governed call runs, the gateway checks it against that policy and records which named rule allowed or denied it. A personal AI account connected straight to work email does not pass through that path, which is the case the guide's section on work warns about.

_Comment from Gil Röder, CPO & Co-Founder_

Source: AEPD, [Asistentes y agentes de IA: Decálogo para personas usuarias](https://www.aepd.es/guias/decalogo-usuarios-asistentes-agentes-ia.pdf), October 27, 2026.
