---
title: AI on staff computers. What may it do? | Disruptive Lab
description: An AI assistant that opens files and acts on a computer needs rights like any colleague. What it may read, what it may do and who decides.
url: https://disruptivelab.dev/en/news/ai-on-staff-computers
language: en
covers: 2026-01-12/2026-01-18
---

1.  [News](https://disruptivelab.dev/en/news)
2.  [AI under the company's control](https://disruptivelab.dev/en/news/ai-under-control)
3.  AI on staff computers. What may it do?

[Guide](https://disruptivelab.dev/en/news/guides)  [AI under the company's control](https://disruptivelab.dev/en/news/ai-under-control)

# AI on staff computers. What may it do?

An AI assistant on a staff computer is only as safe as the rights it is given, so the company, not each user alone, should decide what it may read and what it may do.

- **Published:** 17 November 2026
- **Updated:** 17 November 2026
- **Reading time:** 8 min read
- **Editor:** David Lahuta

Week of 12 January 2026

## In short

AI assistants now work directly on staff computers. Anthropic launched one of them, Claude Cowork, on 12 January 2026, with direct access to local files.[1](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-1) Two days later, by their own account, researchers at PromptArmor showed a hidden instruction in a document making it send a user's file to an outsider's account, with no approval asked at any point.[3](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-3) The vendor's safety page now splits an assistant's tools into tools that read and tools that act, and says the harm depends on both.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) That split is the policy question for every company. Decide centrally what an assistant may read, what it may do, and which actions always wait for a person.

## Contents

1.  [What does an AI assistant do on a staff computer?](https://disruptivelab.dev/en/news/ai-on-staff-computers#what-does-an-ai-assistant-do-on-a-staff-computer)
2.  [How did a hidden instruction make one send files out?](https://disruptivelab.dev/en/news/ai-on-staff-computers#how-did-a-hidden-instruction-make-one-send-files-out)
3.  [Why is this a question of rights, not of a better tool?](https://disruptivelab.dev/en/news/ai-on-staff-computers#why-is-this-a-question-of-rights-not-of-a-better-tool)
4.  [What is the difference between read rights and action rights?](https://disruptivelab.dev/en/news/ai-on-staff-computers#what-is-the-difference-between-read-rights-and-action-rights)
5.  [Who in the company decides?](https://disruptivelab.dev/en/news/ai-on-staff-computers#who-in-the-company-decides)
6.  [Which actions always wait for a person?](https://disruptivelab.dev/en/news/ai-on-staff-computers#which-actions-always-wait-for-a-person)
7.  [What belongs in a company policy?](https://disruptivelab.dev/en/news/ai-on-staff-computers#what-belongs-in-a-company-policy)

## What does an AI assistant do on a staff computer?

It works like a new colleague who has been handed the keyboard. You give it a task, and it carries out the steps by itself instead of only answering a question.

At its launch, Anthropic described Cowork as a research preview for the desktop app that ran in an isolated virtual machine on the user's computer, with direct access to local files and connected tools.[1](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-1) Its safety page today lists what such an assistant does, namely reading files, browsing the web, running programs and using apps.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) The same page says tasks now run in an isolated, temporary environment on the vendor's servers and reach the computer through the desktop app, only for the folders a user has connected.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

Two details matter for a company. Isolation, the vendor writes, limits where the assistant's code runs but not what it reads or does.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) And scheduled tasks run on their own, even when the computer is switched off.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

## How did a hidden instruction make one send files out?

The researchers at PromptArmor describe the case step by step.[3](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-3)

1.  A user connected the assistant to a folder of confidential real estate files, such as valuations and loan estimates.
2.  The user uploaded a document that looked like a set of working instructions for the assistant. Inside it, in white text on a white page at a font size of one point, sat an instruction nobody would see.
3.  The user asked the assistant to analyse the files with those instructions.
4.  The hidden text told the assistant to upload the largest file to the vendor's file service, under the attacker's own account key. The assistant's environment blocked most outside connections, but the vendor's own service counted as trusted.
5.  The file, with financial figures and personal data, reached the attacker's account.

The researchers note that at no point was a person asked to approve anything.[3](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-3) They also write that the underlying weakness had been reported earlier in the vendor's chat product and acknowledged, but not fixed.[3](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-3)

The vendor's answer, as the researchers quote it, was a warning that the product is a research preview with unique risks, and the advice to avoid giving it access to local files with sensitive information.[3](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-3) The same advice stands on the vendor's safety page today, which adds that despite its safeguards the chance of an attack is still not zero.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

This kind of attack has a name. Instructions hidden in content the assistant reads as part of a normal task are called prompt injection.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) The OWASP Top 10 for LLM Applications 2026, the security community's list of the main risks of AI language models, ranks it first and says no reliable way to prevent it exists today.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4)

## Why is this a question of rights, not of a better tool?

Because both the vendor and the security community put the defence in what the assistant is allowed to do. The vendor writes that when something goes wrong, the harm depends almost entirely on what the assistant can read and see and what it may do.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) An attack of this kind succeeds only when two things hold at once, the assistant reads content from outside the sources you trust and it can take actions that could harm you.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

OWASP goes further and advises designing on the assumption that the model's boundary will one day be crossed, and limiting what the model may do so that a successful trick does not become a successful attack.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4) It names three conditions that together make serious harm possible, namely access to private data, untrusted content and a way to send data out, and says that removing any one of them removes the conditions.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4)

In the case above all three were present, a confidential folder, a document from outside and a permitted route out.[3](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-3) A better model would have lowered the odds. Narrower rights would have taken away one of the three.

## What is the difference between read rights and action rights?

The vendor sorts an assistant's tools into two groups.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) Read tools let it see content, for example a mailbox or a screenshot. Write tools let it act, for example create a calendar invitation, delete a file, run a command or click on the screen. The vendor says the second group carries more risk by its nature.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

OWASP traces too much agency to three causes, more functions than the task needs, wider permissions than the task needs and too much freedom to act without a check.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4) Its own example is a mail assistant that only summarises messages. It needs to read mail, not to send or delete it.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4)

In our view, the common rights and who should decide on them look like this.

Right

Typical use

What goes wrong when it is too wide

Who should decide

Read a folder

a project or a department's documents

the assistant sees salaries or contracts it does not need, and every file can carry a hidden instruction

the owner of the data, with IT

Read mail and calendar

summaries, preparing a meeting

an outside sender's message becomes an instruction

IT, per role

Browse the web

research, filling in forms

any page can carry a hidden instruction

IT

Change or delete files

tidying, editing documents

work is lost or overwritten

the user, within limits IT sets

Send, upload or publish

e-mails, shared links, uploads

company data leaves the company

management, enforced by IT

Buy, pay or order

purchases, bookings

money is spent on someone else's instruction

management

Add tools and extensions

new connections to other services

each one opens a new way in

IT

## Who in the company decides?

Not the person at the keyboard alone. The vendor's page is plain that the user remains responsible for every action the assistant takes on their behalf, including messages sent, purchases made and data read or changed.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) Yet many of the settings that decide this sit with the user, such as which folders are connected and how often the assistant asks before acting.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

Some controls do sit with the company. Owners of the vendor's business plans can switch off web search, can send a record of the assistant's activity to their own monitoring tools, and on the enterprise plan can have new extensions scanned for harmful content when installed.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

OWASP adds two principles that make the decision a company one.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4) Actions taken for a user should run with that user's own rights and no more.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4) And whether an action is allowed should be decided by fixed rules in the system, not by the AI model itself.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4)

In our view the work divides simply. Management decides which kinds of action are allowed at all. IT turns that into settings and keeps the record. The owner of each data area decides which folders and mailboxes an assistant may read. The user approves single actions within those limits.

## Which actions always wait for a person?

OWASP recommends that a person approves high impact actions before they happen.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4) It suggests a graded approach, in which easily reversible steps may run on their own while irreversible ones go to a person.[4](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-4)

The vendor's page points the same way. Its assistant always asks before permanently deleting files, whatever mode it runs in.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) It advises switching to manual approval when a task touches sensitive files, uses a new tool or site for the first time, or when mistakes would be hard to undo, such as sending messages or making purchases.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2) In the mode that skips all approvals, it says, nothing checks the assistant's actions.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

For most companies a sensible default list of actions that wait for approval looks like this.

-   anything that sends data outside, such as an e-mail, an upload, a form or a publication
-   deleting or overwriting files and records
-   payments, orders and purchases
-   changes to data in company systems
-   the first use of a new tool, extension or website
-   scheduled tasks, which the vendor advises keeping away from sensitive data and from actions that are hard to undo[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)

## What belongs in a company policy?

A page or two is enough, as long as the settings match it.

-   Which AI assistants may run on company computers, and only on company accounts?
-   Which folders may an assistant read? The vendor suggests a dedicated working folder rather than broad access.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)
-   Which data never goes near it, such as financial documents, access credentials and personal records?[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)
-   Which websites and connected services may it use?
-   Which actions always wait for a person, and does the setting enforce it, not only the policy text?
-   Are scheduled tasks allowed, and for what?
-   Is every step recorded where IT can see it later?
-   What does a user do when the assistant goes beyond the task? The vendor's advice is to stop the task at once and report it.[2](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-2)
-   Who reviews the policy, and how often?

## What this means for a company in Europe

For personal data, the rights an assistant gets are also a legal matter. GDPR asks that by default only the personal data needed for each purpose is processed, and it names the accessibility of that data explicitly.[5](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-5) An assistant connected to a whole disk sees far more than its task needs. GDPR also asks that anyone acting under the company's authority with access to personal data processes it only on the company's instructions.[5](https://disruptivelab.dev/en/news/ai-on-staff-computers#source-5) Rights set centrally, with every step recorded, are a practical way to show that this holds.

## Next step

If it would be useful, we can meet briefly and go through which folders and actions an AI assistant may use in your company, and who approves what. The AI Business Harness works the same way, with each user's permissions and a record of every step, under your company's control. What you do with it afterwards is entirely your decision.

[AI Business Harness](https://disruptivelab.dev/en/ai-business-harness)

## Sources

1.  1. [Anthropic, Claude Help Center, Release notes, entry of 12 January 2026](https://support.claude.com/en/articles/12138966-release-notes). Read on 17 November 2026
2.  2. [Anthropic, Claude Help Center, Use Claude Cowork safely](https://support.claude.com/en/articles/13364135-use-claude-cowork-safely). Read on 17 November 2026
3.  3. [PromptArmor Threat Intelligence, Claude Cowork exfiltrates files (January 2026)](https://www.promptarmor.com/resources/claude-cowork-exfiltrates-files). Read on 17 November 2026
4.  4. [OWASP GenAI Security Project, Top 10 for LLM Applications 2026](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/). Read on 17 November 2026
5.  5. [Regulation (EU) 2016/679, General Data Protection Regulation, Articles 25 and 32, EUR-Lex](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679). Read on 17 November 2026

## Change log

1.  17 November 2026 Published.

## Related articles

Briefing · AI under the company's control

### [Every US giant sells agents. What is real?](https://disruptivelab.dev/en/news/ai-agents-what-is-real)

The large US vendors now sell AI agents. What an agent is in plain words, which tasks suit one today, and what European rules ask of you.

17 November 2026

Guide · AI under the company's control

### [Shadow AI and how to bring it under control](https://disruptivelab.dev/en/news/ai-under-company-control)

Most people who use AI at work use tools the company never chose. What can leave the company, why bans fail and what makes AI the company's own.

17 November 2026

## Company details

- **Company name:** Disruptive Lab s.r.o.
- **Registered office:** Vavrečkova 5262, 760 01 Zlín
- **Company ID (IČO):** 08240647
- **VAT ID (DIČ):** CZ08240647
- **E-mail:** office@disruptivelab.dev
- **Phone:** +420 777 702 032
