# Least privilege projects stall because removing access carries more risk than granting it

Guide, 10 Nov 2026, read 10 Nov 2026

**The short answer.** Cloud least privilege projects stall at four points: nobody can show a removal is safe, usage data covers too short a window, clean-up happens one permission at a time, and the report ends in a ticket. Provider documentation and vendor evidence describe a fix for each of the four.

## Granting is cheap and removal has an owner who fears the outage

AWS documents the pattern in its own IAM guidance: you might start with broad permissions while you explore what a workload needs, and as the use case matures you work to reduce them. The first half of that sentence happens on every project. The second half needs someone to decide that a removal is safe, and that person carries the outage if the decision is wrong.

That asymmetry is the first reason projects stall. A broad grant costs nothing to approve and a failed deploy is visible within minutes. A removal that breaks a quarterly job is visible three months later, with the engineer who approved it as the named cause. Teams respond by freezing the clean-up queue. The fix is evidence per change: a record of what the change would have blocked, before it ships.

## The usage window decides what looks unused

Usage-based tools compare what is granted with what was used in a window. Google documents that the IAM recommender compares a principal's permissions with those used in the last 90 days, and that by default it does not generate recommendations until it has 90 days of data. For project-level recommendations the minimum can be set to 30 or 60 days, and Google says accuracy might be affected. AWS lets policy generation read up to 90 days of CloudTrail events.

A role used once a year for failover, or a job that runs at quarter close, can sit outside any 90-day window. Removing its permissions because they look unused is the exact mistake that makes the next removal harder to approve. Treat the window as a stated limit of the data, and list the identities whose duties run on longer cycles before any removal.

## Permission-by-permission clean-up does not scale

Google's role recommendations each suggest that you remove or replace a role for a principal. That is the right unit for a review and the wrong unit for an estate of hundreds of accounts. If every finding is a ticket, the queue grows at the rate that engineers create roles, which is faster than anyone closes tickets. Role-level review is also slow for a human, since each recommendation needs someone who knows what the principal does.

Some vendors describe a different unit. Sonrai says unused services are locked down in a single global action, and Act says its campaigns remove thousands of access paths with one policy change. Both are vendor claims. The test for either is whether the change can be simulated first and who applies it, which is why this index weights "One change, a class of risk" at 15 percent. See [the method](/method/) for the weights.

## A finding ends the tool's work and starts the team's

AWS Access Analyzer recommends a refined policy for unused permissions and the user attaches it. Google states that the IAM recommender does not apply recommendations automatically, and that you apply or dismiss each one. Both are sound designs for a provider tool, and both leave the removal to a human with a backlog. Neither provider presents this as a gap. A recommendation service recommends, and the remaining work is staffing: someone must read each result, test it and ship it.

This index gives the largest single weight, 25 percent, to whether a product ends with an enforced change or with a finding. [Sonrai Security](/platforms/sonrai/) applies the change through cloud policies the customer deploys. [Act](/platforms/act/) is read-only and delivers each change as a pull request or ticket for the customer to apply. The difference decides who owns the last step. The practical effect is that whoever owns the last step also owns the risk of the change, so the tool's own work is not the whole project.

## New access arrives faster than old access leaves

Even a clean estate decays. Act's engineering blog makes the point that workloads and identities spin up and tear down by the hour, so the map drawn this morning is stale by afternoon. A one-time clean-up is a snapshot, and the next deploy can re-grant what was just removed. Act's post adds that the boundaries to handle this already exist in the cloud, among them security groups, SCPs, RCPs and private endpoints, so the missing piece is keeping them current.

The durable answer is a check before production: a policy test in the pipeline, with exceptions handled in the team's own tools. Act says it blocks risk from shipping in CI/CD. Google documents a separate Terraform plan check against an organization policy library. Whatever the tool, a project that has no check before production is cleaning a floor while the tap runs.

## Sources

- AWS IAM best practices, apply least-privilege permissions: https://docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html (read 10 Nov 2026)
- Google Cloud, overview of role recommendations: https://cloud.google.com/policy-intelligence/docs/role-recommendations-overview (read 10 Nov 2026)
- AWS IAM Access Analyzer policy generation: https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-policy-generation.html (read 10 Nov 2026)
- AWS, what is IAM Access Analyzer: https://docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html (read 10 Nov 2026)
- Act blog, network and identity each see half your cloud: https://act.security/resources/network-and-identity-each-see-half-your-cloud-your-attacker-sees-both (read 10 Nov 2026)
- Sonrai Security, least privilege use case: https://sonraisecurity.com/use-cases/least-privilege (read 10 Nov 2026)
- Google Cloud, validate Terraform plans against policies: https://cloud.google.com/docs/terraform/policy-validation/validate-policies (read 10 Nov 2026)
- Act Security, platform page: https://act.security/platform (read 10 Nov 2026)

Last reviewed 10 Nov 2026. Canonical: https://cloudleastprivilege.com/guides/why-least-privilege-projects-stall/
