Write Permissions

On a computer, it is precisely defined who is allowed to touch which data. This definition is called rights or permissions. There are two very different levels here. Whoever has read rights may view a file, but not change anything about it. Whoever has write-permissions, meaning write rights, may overwrite, add to, or completely delete the content. The difference is roughly like that between a book in a library and a book you’re allowed to write in with a pen.

Why writing is more dangerous than reading

Reading cannot be undone, but it also destroys nothing. Writing, on the other hand, changes a state. An incorrect write operation can overwrite data that nobody can ever retrieve again. Precisely for this reason, experts treat write rights as the significantly more delicate right.

A second point is the chain reaction. Whoever is allowed to write in a database can also alter entries about other rights there. A small write right can thus, under certain circumstances, turn into control over an entire system. Attackers therefore specifically look for places where write rights have been distributed too generously.

For this reason, companies follow the principle of least privilege. Every program and every employee receives only exactly the rights necessary for the task. An analytics program that displays sales figures doesn’t need write access to the sales data. It’s enough if it’s allowed to read.

How rights are assigned technically

For files, the operating system stores for each file who may read, write, and execute it. A distinction is usually made between the owner, a group, and everyone else. On Linux systems, this can be seen in abbreviations like rw-r--r--. This means: the owner may read and write, everyone else may only read.

With online services, it works via so-called tokens. A token is a digital key that a program presents in order to gain access. This key contains information about what is permitted. A token with read scope may only query, a token with write scope may also make changes. Anyone who has ever granted an app access to their account has confirmed exactly such a list.

An important point is often overlooked. Write rights are not re-checked with every single operation, but remain valid for as long as the key is valid. A stolen token with write rights is therefore just as effective as the real one. Good systems therefore issue write tokens with only short validity and log every write operation along with the timestamp.

AI agents and the question of how much they’re allowed to touch

In news about artificial intelligence, the term has appeared since the emergence of so-called agents. These are AI programs that don’t just respond, but carry out tasks themselves. Such an agent is supposed to, for example, fix program code in a project or enter appointments in a calendar. Both are impossible without write-permissions.

This is exactly where the point of contention lies. A language model makes decisions based on probabilities and occasionally gets it wrong. It can also be deceived by hidden instructions in a text. An agent with full write rights could, in such a case, delete real data. Providers therefore work with intermediate steps: the agent proposes a change, a human confirms it.

In everyday life, one encounters this topic more often than one might think. A shared document in a cloud can be shared as “view only” or as “can edit”. On developer platforms like GitHub, write access determines who may directly contribute changes to a project. And when an app asks for permission for photos, the question is whether it may only view them or also delete them.