Claude Cowork limits and safety. Files from your disk are processed on Anthropic's servers
How Cowork uses your limit, where your files go, what prompt injection is, which approval mode to choose and what not to give Claude Cowork.
Part of the guideClaude Code, the complete guide. From installation to your own skills and plugins
In this article
In short
Cowork uses your plan's limit faster than ordinary questions, because every step of a task costs tokens, and you can check usage in Settings > Usage. Tasks run on Anthropic's servers, so files Claude opens from your disk are processed outside your computer. The biggest risk is prompt injection, instructions hidden in content Claude reads during an ordinary task.
A business owner gives Claude Cowork a folder of client emails and asks for a summary. In one of the emails someone added a line in small print that pretends to be an instruction for the AI. Anthropic describes exactly this kind of example and says plainly that the risk of an attack is not zero. So before Cowork tasks become routine, it is worth knowing how its limits and safety work.
This article is part of our Claude Code guide. It is based on Anthropic's help centre, mainly the article on using Cowork safely, as of 10 November 2026. It describes what the vendor states, not a legal assessment.
How Cowork uses your limit
Everything you do with Claude counts towards your plan's limits. Multi-step tasks, ones that run code, create files or use connected apps and the browser, use more than a quick question. Anthropic puts it simply. Each step Claude takes uses tokens.
On paid plans you see it in Settings > Usage. It shows the five-hour session bar with the time until it resets, and the weekly limits. On the Ultra plan verified accounts have no five-hour session limit.
Anthropic gives a few ways to save your limit in Cowork. Group related work into one task instead of several. Start a new conversation for unrelated work, so Claude does not carry extra context. When you only need a quick answer, not a file, say so. More on the limit is in why the Claude usage limit runs out fast, and on tokens in tokens and the context window.
Where your files go
Cowork tasks run in the cloud. For each session Anthropic creates a separate, temporary environment on its servers that cannot reach your home or company network and is removed when the session ends. When a task needs a file from your disk, Claude reaches it through the Claude Desktop app, only in connected folders and only while the app is open.
One sentence in Anthropic's help matters most here. Claude's work, including local files it opens through the app, is processed on Anthropic's servers rather than staying on your computer. Claude fetches a copy of just the file it needs. When you delete the session, those copies are deleted too.
A deleted task disappears from your history at once and from Anthropic's systems within 30 days. Whether your conversations help train models depends on the Help improve our AI models setting in Settings > Privacy. On the Enterprise plan admins can retrieve Cowork sessions through the Compliance API. Local sessions on staff computers in Team and Enterprise plans keep their history on those computers.
Whether this processing is fine for your clients' data is a question for a lawyer, not for us or for an AI. A short company AI policy helps, and client data in chats is covered in client data in ChatGPT.
Prompt injection, instructions hidden in content
Anthropic uses this name for an attack where someone hides instructions in content Claude reads during an ordinary task. In the example from its help, Claude summarises emails, and one of them contains a line telling it to ignore earlier instructions and transfer money to a given account. A successful attack makes Claude follow the attacker's instructions instead of yours.
For such an attack to work, two things must be true at the same time. Claude must read content from outside what you consider trusted, such as websites, emails from strangers or downloaded files. And it must be able to do something that harms you. If one of these is missing, the attack becomes much harder.
Anthropic trains Claude to recognise such instructions, scans untrusted content with classifiers and, in Auto mode, checks each action before it runs. Even so, it writes that the chance of an attack is not zero. If Claude suddenly starts talking about something unrelated, reaches for files or sites you never mentioned, or asks for sensitive data on its own, stop the task and report it with the feedback button in the app, or through your admin on Team and Enterprise plans.
Approval modes and deleting files
Anthropic splits Claude's tools into two groups. Read tools only take in content, for example reading your inbox. Write tools change something, for example create a calendar invite, delete a file or click on the screen. The second group carries more risk, so Cowork treats it with more care.
How often Claude asks for permission is set by the mode in the message box. Manual means asking before each action. Auto means working without asking, with an automatic check of each step. Skip, in the old view, means nothing checks Claude's actions. Anthropic advises switching to Manual when a task touches sensitive files, accounts or sites, when you use a new tool or plugin for the first time, and when mistakes would be hard to undo, such as sending messages or making purchases.
In every mode Claude asks for explicit permission before it permanently deletes a file.
What not to give Claude Cowork
Anthropic's list is longer, but these points come up most often.
- Files with sensitive data, such as financial documents, credentials and personal records. Rather than giving access to the whole disk, create a dedicated working folder and keep backups of important files.
- Sites where you are signed in that handle money or personal data. This applies to the built-in browser and to Claude in Chrome.
- Scheduled tasks that send messages on your behalf, make purchases or touch sensitive files. Nobody watches them while they run.
- Unfamiliar plugins and MCP servers. Local MCP servers run on your computer with the same permissions as any other program.
- Computer use around banking and health apps. There is no sandbox between Claude and your screen, and Claude takes screenshots to see what is on it.
Anthropic also reminds you that you are responsible for everything Claude does on your behalf. Messages sent, payments, changed data and actions taken by scheduled tasks. That includes respecting the terms of the sites Claude visits.
To start, keep a simple setup. A folder of copies, Manual mode for anything that leaves the company, and no passwords in connected folders. How to install Cowork is in Claude Cowork setup, and ready tasks to start with are in Claude Cowork for office work.
If you want to know where AI can take work off your company, ask for our free AI audit. We look at your tasks and tell you where to start.
What would you like to change in your business?
Leave your number. We will talk about your process and see whether we can help. The first call is free.
The service this text is about
Automation and AI / OCRAutomation of document re-typing, OCR and checking rules, and your data stays with you.