START HERE
What you will learn
- Separate network reachability, identity and permission.
- Define the smallest useful authority for an integration or agent.
- Explain how independent safeguards address different failure paths.
Before you start: The SSH and tool-permission examples are easier to follow if you understand an account, a resource and an operation.
Reference edition: NIST SP 800-207 and OWASP authorisation guidance. Sources checked 2026-11-11.
Being connected is not the same as being permitted
A VPN route may let a device reach a service, while authentication establishes an identity and authorisation decides whether it may perform a particular operation. Treat these as separate questions. An internal network address is not, by itself, evidence that every requested action should be allowed.
NIST’s zero-trust architecture guidance focuses on resources and identities rather than implicit trust from location or ownership. That is an architectural approach, not a product label or a promise that every request is safe. The actual policy still needs to specify what may happen and under which conditions.
Use multiple safeguards for different jobs
Least privilege asks which resources and operations an identity actually needs. A release identity might activate a prepared website version without needing general administrator authority. A research assistant might read approved documents without needing a send-anywhere email tool.
Defence in depth means useful layers, not several products repeating the same assumption. A private management route limits reachability; SSH host verification checks the destination; account permissions limit operations; release checks inspect the outcome. For a sensitive operation, an unavailable permission decision should not silently become an approval.
Step-by-step process
List resources and operations
Identify what can be read, written, deleted, published or executed. Name the consequences rather than calling every capability “access”.
Assign identities to bounded jobs
Separate routine release work from host administration, and separate reading from changing data where the workflow allows it. Avoid reusing a broad credential just because it is available.
Make permission decisions at the operation
Check the caller, target and allowed operation in the enforcing service. The interface or model’s proposed action is not the final authority. Any separate approval must refer to the specific target and effect being requested.
Design for a missing or failed check
Specify what happens when identity, policy or approval information cannot be established. For a protected operation, stop or escalate instead of granting new authority by default.
Check the layers and revocation path
Use authorised tests to confirm denied operations, credential removal and recovery decisions. Inspect whether two apparent safeguards depend on the same unchecked assumption.
FOLLOW ONE EXAMPLE
Worked scenario
A deployment assistant requests an unnecessary administrative action
A fictional release assistant needs to upload a prepared archive and activate it. It does not need to modify account permissions on the VM.
- The release identity is allowed to perform its defined release operations but not account administration. A separate management path exists for authorised host changes.
- If the assistant proposes changing a user’s permissions, the enforcing account denies the operation even if the connection arrived through the private route.
- The operator inspects the request and the release result. They do not broaden the release identity merely to make an unrelated command succeed.
What this shows: The network route, verified connection and narrow account permissions have different jobs. The denied operation keeps an unnecessary capability outside the release path.
The process at a glance
Read Mermaid diagram source
flowchart TD
A[Request reaches the service] --> B[Establish the identity]
B --> C{This operation is permitted?}
C -->|No| D[Deny and record]
C -->|Yes| E{Separate approval required?}
E -->|Yes| F[Human reviews target and effect]
E -->|No| G[Perform the bounded operation]
F --> I{Approval granted?}
I -->|Yes| G
I -->|No| D
G --> H[Check the result]What you should have at the end
- A resource-and-operation view of the permissions needed.
- Separate identities or permissions for jobs with different consequences.
- Evidence that disallowed operations remain denied and access can be revoked.
Common mistakes and how to avoid them
Treating a VPN address as complete authorisation
A permitted route does not decide every resource operation.
Calling many overlapping tools defence in depth
Ask which distinct failure each layer would actually address.
Expanding permissions to resolve an unexplained failure
Investigate whether the operation belongs in the task before granting more authority.
APPLY THE IDEA
Try it yourself
Your task: List five capabilities of one integration you own. Mark each as required, unnecessary or requiring a separate approval.
What to produce: A proposed least-privilege permission set and one authorised negative test for an operation it should deny.
Check your understanding: If the model or operator requested more than the task needs, which component would enforce the limit?
Key terms in plain English
- Least privilege
- Giving an identity only the authority needed for its defined work.
- Zero trust
- Avoiding implicit trust based only on location or asset ownership.
- Defence in depth
- Using distinct safeguards to reduce different paths to a harmful outcome.
- Deny by default
- Rejecting an operation unless the required permission is established.
- Revocation
- Removing previously granted access, such as deactivating a credential.