Least Privilege

Least privilege is the principle that any component should hold only the permissions it needs to do its job. Its value is not that it prevents compromise but that it bounds it: the question after any breach is what the compromised credential could reach.

The principle is stated easily and applied poorly, because broad permissions are convenient and the cost of narrowing them is paid up front. A few recurring patterns:

Scope tokens to operations, not to roles. A process that only reads should hold a read-only credential. A process that only posts a comment should not hold one that can write to the repository.

Know which credential is actually in play. A permissions declaration frequently bounds the platform-issued token while some other, separately-minted credential does the real work — so tightening the declaration achieves nothing. See Environment Variable for the related problem of an unexpected credential taking precedence.

Constrain outbound capability, not just inbound. For anything running untrusted input through an agent, the exfiltration path matters more than the permission set. See Prompt Injection and Supply Chain Security.

Wildcards are not a configuration. An allowlist of "anything" in a system that defaults to nothing is a common and understandable shortcut, and it grants exactly what the mechanism existed to restrict.