Vibeleaderboard
← Course overviewModule 02 · Understand the software underneath · Beta
Lesson 2.5

Module contents

Understand the software underneath

Lesson 5 of 6

Who is allowed to read or change it?

The garden board accepts requests over the network, but only some requests should be able to publish a post. This lesson separates proving an identity from permission to act, then explains how software can enforce a boundary that a conversation alone cannot.

Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.

Identity, proof and permission are three different things

An account is a record a service keeps about a person or a program: an identity the service recognizes. A is something presented with a request to prove that the sender is that identity: a password, a code from a phone, a long generated key that a program carries. The check that a credential matches the identity is : is this who it claims to be? A separate check decides what that identity may do. That is : is this identity permitted to perform this operation on this resource?

MDN's guide to authentication on the web describes the mechanism: a client includes credentials in its request, the server verifies them, and even a request with valid credentials can be refused when that identity lacks permission for the resource. In the authentication flow it describes, a "forbidden" status can signal valid credentials with insufficient permission. A status alone does not always reveal the exact cause; services can also conceal a resource by reporting it as not found. 7 The two refusals mean different things. "Who are you?" and "you may not do that" are separate answers, and a system can give either.

Apply it to the board. The organizer's client sends the post with the organizer's credential. The server authenticates: this credential belongs to the organizer's account. Then it authorizes: the organizer's account may post to the Saturday announcements. Both pass, and the post is stored. A volunteer's account might authenticate perfectly and still be refused, because it is allowed to read announcements and not to post them. The volunteer is not logged out. The volunteer is logged in and not permitted. An application does not necessarily receive every permission you have. A service can grant it access to selected resources or operations, and the application can apply further restrictions. GitHub, a service for storing and collaborating on software projects, documents one example: an app acting for a user can access only what both the user and the app are permitted to access. 14 For the garden application, you need to know which account it uses, what access was granted to it, and which additional limits are enforced.

Keep passwords and secret keys out of the conversation

Passwords and secret access keys can let someone act using the access they unlock. Copying them into a conversation or shared document can expose that access. Other login methods require additional proof or a particular device; the consequence depends on the method, so “every credential grants the whole account” would be the wrong rule. 7

Module 1 explained that an application decides what enters the model's and what conversation material it stores. A secret pasted into a message may enter both. Use the application's documented account-connection or secret-entry mechanism instead of putting secret values in the conversation. Check what access the connection grants. Later modules explain these connections in more detail.

An instruction is not a restriction

Now the distinction this module most needs to leave behind, because module 3 depends on it. Suppose you tell an assistant application, in the conversation, "only read files in the Garden folder, and do not send anything." That sentence is context. It goes to the model, which may follow it; module 1 explained why following is likely and not guaranteed. The sentence alone does not revoke permissions. Suppose our fictional application has tools that can read the Garden folder and publish to the board, and no separate rule blocks publishing. The model may follow the request, but the posting capability remains available. Another application might enforce additional limits; that is something to establish from its controls, not infer from the conversation.

A , in the sense this course uses from now on, is a limit enforced by software other than the model: the operating system refusing a process access to files outside a named folder, a network rule that allows connections only to listed destinations, a credential that is authorized to read and not to post. A correctly configured restriction blocks the covered operation even if the model requests it. Its protection depends on what it covers and whether the enforcement works. Anthropic, describing how it contains its own Claude products, gives examples of exactly this separation: boundaries on file access and network destinations enforced outside the model, with different designs for different products. It also describes ways such containment can fail, including allowed destinations that were drawn too broadly. That is a vendor's account of its own systems. It is useful here as an illustration that the boundary is a property of the surrounding software and has to be examined product by product, not as a description of how every application behaves. 9

So when an application says it only has access to something, the question is which software enforces that and what it covers. An instruction in the conversation is a request for behavior. An enforced restriction blocks specified actions within its scope; it is not a guarantee against every possible failure. Module 3 builds its explanation of permissions on this difference.

Carry this forward

  • An account is an identity a service recognizes; a credential proves it; authentication checks the proof; authorization checks the permission. Valid credentials can still be refused.
  • Check the access granted to the application and any additional restrictions. Acting for you does not necessarily give it every permission you have.
  • Passwords and secret access keys can expose the access they unlock. Use documented account connections and secret-entry mechanisms, not conversation text.
  • An instruction in the conversation asks the model to behave. A restriction is a limit enforced by other software. Ask which software enforces a claimed boundary and what it covers.

Lesson check

A fictional assistant has working tools and granted credentials to read the Garden folder and publish to the board. No separate software rule blocks publishing. The organizer types: “Read the folder and draft a summary. Do not post anything.” Which statement follows?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

The post is allowed and it succeeded. Whether it did what you asked is still a separate question, and the last lesson makes checking it a habit.

Sources · Reviewed 2026-09-28
  1. 7. MDN: HTTP authentication
  2. 9. Anthropic engineering: How we contain Claude across products
  3. 14. GitHub Docs: permissions for apps acting on your behalf

Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.