Zum Hauptinhalt springen
Misuse pattern · AI Act

MU-11 Passing on access credentials

Updated on 6 min Reviewed by: Cenedril Editorial
EU AI ActISO/IEC 42001

An engineering firm has an enterprise account for an AI assistant with a data processing agreement and access to the project files. A project manager gives his login to an external subcontractor so that he can get to the documents faster. The subcontractor soon uses the assistant for other clients’ projects too and uploads their plans. A vetted tool with a known group of users has turned into an open account whose users nobody has an overview of any more.

This pattern describes access that is passed on to people or organizations outside the intended group of users. For the AI Act (Regulation (EU) 2024/1689), it is reasonably foreseeable misuse under Article 3(13), because shared logins are among the oldest habits in dealing with software.

What is behind it?

The approval of an AI system applies to a specific group of users. Contract, briefing, permissions and logs all rest on the assumption that it is known who uses the system. An account that is passed on overturns that assumption in one stroke. With AI assistants that can reach mailboxes, file storage or customer data, it also opens the way to everything the assistant can see.

How it happens in daily work

  • Licence costs. There are fewer seats than users, so accounts get shared.
  • Working with externals. Service providers, partners or customers are supposed to join in quickly, and a regular guest account takes too long.
  • Group accounts. A team shares one account because that seems easier.
  • Private co-use. The company account is used by others at home as well.
  • Technical keys. API keys end up in source code, chat messages or shared documents and thus with anyone who finds them.

Why it is risky

Whoever uses someone else’s account has not been briefed and is not bound by the organization’s rules. Data the assistant can reach becomes visible to outsiders. Logs can no longer be attributed to a person, so neither oversight nor the investigation of an incident is possible. If an external organization uses the account for its own purposes, it uses the system under its own authority and is thus itself a deployer under Article 3(4) of the AI Act, without knowing it. Add breaches of the licence terms, which can go as far as the account being blocked.

Impact

The impact ranges from unexpected costs to breaches of confidentiality and content generated in the organization’s name without its knowledge. If the provider blocks the account, an entire workflow may come to a halt in the worst case. Whoever passes on an account passes on everything that account can do.

Practical examples

Group account in customer service. Ten customer service staff share one account for an AI reply system. When an insulting reply goes out to a customer, it cannot be established who approved it. The investigation ends without a result, and the cause remains.

API key in a public repository. A developer uploads a project with an embedded API key to a public code repository. Strangers use the key for weeks for their own requests. The provider’s bill multiplies, and because content was created through the account that breaches the provider’s terms, the provider blocks access.

Agency account at the client. An advertising agency gives a client the login to its AI image generator. The client uses it to create realistic images of a well-known person for its own campaign and publishes them without any label. With the provider, the breach falls back on the agency whose account was used.

Effective controls

The following ISO/IEC 42001 controls counter this pattern.

Prevention:

  • A.10 Third-party and customer relationships: Responsibilities between the organization, its partners, suppliers and customers are clearly allocated. Anyone who gives externals access settles use, data and confidentiality beforehand, and the provider’s terms are respected.
  • A.9 Use of AI systems: The processes for responsible use define who may use a system and under which conditions.
  • A.6.2 AI system life cycle: The deployment requirements include access management that attributes actions to individual people.

Detection:

  • A.6.2 AI system life cycle: Event logs show simultaneous logins from different places, unusual usage volumes or access outside working hours.

Response:

  • A.9 Use of AI systems: Shared accounts are blocked, keys replaced, the reachable data assessed and the people affected informed where necessary.

The basics come from the ISO/IEC 27001 controls on authentication information (A.5.17) and access rights (A.5.18) and from the Starter Kit’s access control policy. The AI Usage Policy expressly includes in Section 2 external staff who use AI on behalf of the organization, and requires in Section 7 that every suspicion of inadmissible use is reported.

For high-risk AI systems, the provider evaluates foreseeable misuse under Article 9(2)(b) and names circumstances that may lead to risks in the instructions for use (Article 13(3)(b)(iii)). The deployer uses the system in line with those instructions under Article 26(1) and assigns human oversight under paragraph 2 only to people with competence, training and authority. Both assume that it is known who operates the system. The logs the deployer keeps under paragraph 6 and the record-keeping duties under Article 12 only help if actions can be attributed to people (→ EU AI Act).

Who counts as a deployer is set by Article 3(4): any person or body using an AI system under its authority. An organization that uses a passed-on account for its own purposes falls within that. Article 4 also applies to everyone: whoever lets externals work with AI on its behalf also supports their AI literacy.

Sources

Frequently asked questions

Is a shared team account for an AI tool acceptable?

It often already breaches the provider's licence terms. Above all, it makes logs worthless, because actions can no longer be attributed to a person, and it prevents briefing and permissions from matching the individual. Personal accounts with single sign-on are almost always the better choice.

How do I give external people safe access to an AI system?

Through a separate, time-limited guest account with the permissions the task requires and contractual rules on use, data and confidentiality. The external person is briefed on the rules before first access. When the assignment ends, the account expires automatically.

What do I do if an API key has become public?

Revoke and replace the key at once, then check usage for the period in question with the provider: which requests ran, which data was reachable, what costs arose? If personal data is affected, the data protection assessment follows. Afterwards the key belongs in a secrets vault.

No Cookies!

This wiki collects nothing, bakes nothing and leaves nothing behind. There's nothing to consent to. Privacy doesn't get better than this.