Yes, all three clouds document "check usage before you tighten" for service identities - but only Google's replays real access against the proposed rule, and each one has a window or coverage hole a scheduled job can fall through
The question
Verbatim: "Does any major IAM vendor document a "check what was actually used before you tighten" workflow that covers non-interactive service identities, not just human users?"
Open follow-up #1 of [[2026-11-06-correct-rule-wrong-default-permission-detection]], which named AWS IAM last-accessed data, IAM Access Analyzer unused-access findings, GCP Policy Simulator and the Azure/Entra equivalents as candidates but had spent its caps and labeled all of them unverified. This is the verification pass.
What we already know (from the vault)
- The parent brief found no source that names "correct rule, wrong default" as a failure class, and split the available controls into two named families: alert-on-absence after the fact, enforcement staging before the fact. It verified only two staging mechanisms (OPA Gatekeeper
dryrun/warn, Entra report-only mode for Conditional Access) and observed that both assume an administrator distinct from the affected workload ([[2026-11-06-correct-rule-wrong-default-permission-detection]]). - The parent's sharpest finding was that staging is bounded by window length versus job period: a job that does not fire inside the observation window is invisible to it. That was sourced to a single community blog and flagged as unverified against vendor docs. This brief can now partly upgrade it (see below).
- RDCO's own instance: 251 denial events across 401 cron runs, all 401 reporting
subtype: "success", with two state files frozen since 2026-08-17 ([[2026-10-23-rdco-auto-mode-classifier-miss-profile]]). The denying rule is a harness-level protected-path check with no dry-run mode and no allow-rule override ([[2026-10-24-coded-vs-prompted-agent-permission-gates]]). - RDCO has one prior case of auditing a permission against what actually exists rather than what a document says exists: [[2026-07-16-fat-cat-board-router-token-scope]] found an audit item had been "inherited from a proposal, not from production." Cited from its title and summary only; it is the same instinct this brief asks about, applied by hand.
- [[2026-07-25-agent-first-class-identity]] holds the vault's position that session-level identity (a shared credential, a persona) is insufficient for agents. Relevant here because every vendor tool below keys its usage data to a distinct, persistent identity; without one there is nothing to attribute usage to.
What the web says
Three vendor pages were fetched and read on 2026-11-19 (AWS, Google, Microsoft). Claims marked search-level come from search-result summaries of vendor doc pages that were not themselves fetched.
| Candidate | Covers service identities? | Lookback | Pre-enforcement or post-hoc | Verification |
|---|---|---|---|---|
| AWS IAM last accessed information | Yes: roles, by name | Service level "at least 400 days"; action level from a per-service start date | Evidence to read before a change; does not test a proposed policy | Fetched |
| AWS IAM Access Analyzer unused access | Yes: unused roles and unused permissions; service-linked roles excluded | Configurable tracking period, 1 to 365 days | Continuous findings; does not test a proposed policy | Search-level |
| GCP Policy Simulator | Principal-agnostic by mechanism (replays access logs); the overview page does not enumerate principal types | 90 days | True pre-enforcement: replays real access attempts against the proposed policy | Fetched |
| GCP IAM recommender / service account insights | Yes: docs use a service account as the worked example | 90 days max; 30 or 60 selectable for project-level | Recommendation from past usage | Search-level |
| Entra Conditional Access for workload identities, report-only | Partly: single-tenant service principals only; managed identities excluded | None backward; observes sign-ins after the policy is saved | Pre-enforcement, forward-looking staging | Fetched |
- AWS documents the workflow in so many words, and names roles. The last-accessed page opens: "As an administrator, you might grant permissions to IAM resources (roles, users, user groups, or policies) beyond what they require. IAM provides last accessed information to help you identify unused permissions so that you can remove them." Its caveat section begins: "Before you use last accessed information from a report to change the permissions for an IAM identity or AWS Organizations entity, review the following details about the information." That sentence is the closest thing found to a vendor literally writing "check what was used before you tighten." The page never uses the words "service account" or "workload"; roles are simply how AWS workloads authenticate, so coverage is by identity type, not by an explicit non-human callout (docs.aws.amazon.com).
- AWS's lookback is the longest found, but the detail runs out exactly where a scheduled job lives. "The tracking period for service information is at least 400 days," long enough to see an annual job. Action-level detail is narrower: "Action last accessed information is not available for any data plane event," and "The
iam:PassRoleaction is not tracked." So the data can say a role touched S3 in the last 400 days but not that it wrote objects, which is the question a tightening actually turns on. Two more stated limits: the data "includes all attempts to access an AWS API, not just the successful attempts" (an already-denied job still reads as "in use"), and access granted through resource-based policies, SCPs, permissions boundaries and session policies "is not included in your report" (same page). - Access Analyzer unused access is the continuous version, with a configurable window. Search-level: the analyzer flags roles inactive for the configured number of days plus unused permissions, with a tracking period of 1 to 365 days, and does not analyze service-linked roles. A Well-Architected lens page warns that rarely exercised permissions (create, delete) can look unused in a short window. One AWS SDK reference gives the range as 1 to 180 days; unresolved (access-analyzer-create-unused, well-architected lens). The fetched last-accessed page confirms the link between the two: "You can continuously monitor last accessed information with unused access analyzers."
- Google's Policy Simulator is the only true pre-enforcement test found. It "lets you see how a change to an allow policy might affect a principal's access before you commit to making the change," and "You can use Policy Simulator to ensure that the changes you're making won't cause a principal to lose access that they need." Mechanism: it "determines which access attempts from the last 90 days have different results" under the proposed and current policy. Quotes are from a fetch-model extraction of the page, not my own read of the full text (docs.cloud.google.com). Search-level: a separate simulator exists for deny policies, also with a 90-day replay.
- The simulator's limits are documented and they are the limits that matter. Lookback is fixed at 90 days. "The maximum number of access logs that a simulation can replay is 5,000," and with more than that "the simulation will fail" (a search summary of the gcloud reference says 1,000; unresolved). It "does not support all resource types," returns "unknown" when it "did not have enough information to fully evaluate the access attempt," and tells the reader "It's important to review these errors so that you understand the potential gaps in the simulation." For credential access boundaries the page states the window hole outright: "if the token broker's access isn't used within 90 days, their access isn't included in the simulation," and in that case "Policy Simulator fails to report a loss of access."
- Google's recommender covers service accounts explicitly. Search-level: role recommendations use at most the most recent 90 days of permission usage; the docs' worked example is a service account that has not used any Compute Admin permission in 90 days; service account insights flag accounts that have not authenticated in 90 days; the minimum observation period can be lowered to 30 or 60 days with a stated accuracy warning (role-recommendations-overview, service-account-insights).
- Microsoft covers service identities for staging, not for usage history, and with large exclusions. Conditional Access for workload identities can be "saved in Report-only mode, allowing administrators to estimate the effects," with results in the service principal sign-in logs. But: "Policy can be applied to single tenant service principals that are registered in your tenant. Microsoft and third-party SaaS applications, including multitenant apps, are not covered by these policies. Managed identities aren't covered by policy." The only grant control is block, the conditions are location and service principal risk, policies assigned via a group "are not enforced for that service principal," and editing requires Workload Identities Premium licenses (learn.microsoft.com). No Entra feature that reports which permissions a service principal actually used was verified this round; that is a gap in this brief, not a finding that none exists.
Convergences and contradictions
- The parent's "every staging mechanism assumes a separate administrator" finding holds across all five. AWS addresses "an administrator," Google's simulator is run by whoever proposes the policy, and Entra says report-only is for "administrators to estimate the effects." None documents the workload checking its own impending restriction, and none addresses operator-equals-administrator.
- The parent's window-versus-period claim is upgraded from community blog to vendor documentation, for Google. The simulator page states that access not used within 90 days "isn't included in the simulation" and that the tool then "fails to report a loss of access." That is scoped to credential access boundaries in the text, but the mechanism (replay only what happened in the window) is general. My inference: the same hole applies to any job with a period over 90 days. AWS's 400-day service-level window is the only one that would see an annual job.
- Contradiction with the parent's framing that pre-enforcement tooling "evaluates the policy against live traffic." That is true of Gatekeeper and Entra report-only, which observe forward. Google's simulator and both AWS tools look backward at recorded usage instead. These are two different staging designs: forward observation costs a waiting period equal to the longest job period; backward replay costs nothing in waiting but is only as good as the log retention. The parent brief treated them as one family.
Synthesis for RDCO
The answer is yes, with a ranking. The four named candidates are all real and all cover non-human identities to some degree. Ranked by how close each is to a pre-enforcement control for the correct-rule-wrong-default failure: (1) GCP Policy Simulator, the only one that takes a proposed rule and replays actual recorded access against it; (2) AWS last-accessed data and Access Analyzer unused access, which document "review usage before you change permissions" explicitly and cover roles, but report evidence rather than testing a specific change; (3) Entra report-only for workload identities, which is forward staging for a narrow policy type and excludes managed identities and multitenant apps. Confidence is high on the fetched rows (AWS last-accessed, Policy Simulator, Entra) and medium on the two search-level rows.
The useful correction to RDCO's thinking is that staging comes in two designs and RDCO only wrote down the slow one. The parent brief's control (3) says: before a permission change, hold it for one full period of the longest-period affected job, then check mtimes. That is forward observation, the Entra design, and it is why a change touching a monthly job "cannot be judged for a month." Google's design is the alternative: do not wait, replay. Take the proposed rule and run it against the last N days of what the workloads actually did. For RDCO this is an inference, not something any vendor documents for agent harnesses, but it is the transferable part: the raw material for a replay is a record of each cron run's tool calls and paths, and the lookback has to be at least as long as the longest job period. Google picked 90 days; RDCO's longest standing periods are monthly (finance-pulse, acquisition-rescan), so roughly 35 to 45 days of retained run records would cover every current slug. Whether the cron runner retains enough per-run detail to replay against is not something this brief checked.
Every vendor documents the holes, and the holes line up with RDCO's failure. AWS: no data-plane actions at action level, so the tool knows a role used a service but not that it wrote. RDCO's freeze was a write failure with reads intact. AWS again: denied attempts count as usage, so a job already failing still looks "in use" - worth knowing as a trap in any homegrown replay that counts attempts rather than successes. Google: anything outside the window is silently absent and the tool "fails to report a loss of access." Entra: the identity types most used for unattended automation in Azure (managed identities) are outside the policy entirely. The pattern is that these tools are strongest on interactive, frequent, control-plane access and weakest on infrequent, scheduled, data-plane access, which is the profile of a nightly or monthly agent job. So the industry's best pre-enforcement control is real but would not have been sufficient alone, and the parent brief's post-hoc freshness manifest stays necessary rather than becoming redundant.
For the phData-facing side, this is a clean, citable three-cloud comparison for any client conversation about putting agents into production under least privilege: the tooling to right-size a service identity exists on all three clouds, it is usage-window-bound, and an agent that runs monthly or does mostly data-plane work is where it under-reports. That is a defensible, vendor-sourced caution rather than an opinion.
Why this is in the vault
It closes follow-up #1 of [[2026-11-06-correct-rule-wrong-default-permission-detection]] by converting four named-but-unverified candidates into verified or search-level rows, and it adds the backward-replay staging design as an alternative to the hold-for-one-period step that brief specified for RDCO's own permission changes.
Open follow-ups
- Does Microsoft document any usage-history view for service principal permissions (which granted Graph or Azure RBAC permissions a workload identity actually exercised)? This round verified only the Conditional Access staging path; the Entra row for "what was used" is empty, not negative.
- Do AWS IAM Access Analyzer custom policy checks (
check-no-new-access,check-access-not-granted) or the IAM policy simulator give AWS a proposed-policy test comparable to GCP's replay? Named in the parent brief and still unverified; if yes, the ranking above changes. - Is the 90-day fixed window in Policy Simulator and the recommender documented as a design choice with guidance for longer-period workloads, or just a retention limit? Google documents the hole for credential access boundaries; whether it gives guidance for quarterly or annual jobs was not found.
- Does the Claude Code headless result or transcript retain per-run tool-call and path detail sufficient to replay a proposed permission rule against past runs? Overlaps the parent's follow-up on richer success signals; unanswered in the research dir.
Related
- [[2026-11-06-correct-rule-wrong-default-permission-detection]]
- [[2026-10-24-coded-vs-prompted-agent-permission-gates]]
- [[2026-10-23-rdco-auto-mode-classifier-miss-profile]]
- [[2026-07-16-fat-cat-board-router-token-scope]]
- [[2026-07-25-agent-first-class-identity]]
Sources
Web, primary, fetched 2026-11-19:
- https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_last-accessed.html - identity types covered, "Before you use last accessed information..." caveat list, 400-day service tracking period, no data-plane action data, attempts-not-successes, excluded policy types, pointer to unused access analyzers. Full page text read.
- https://docs.cloud.google.com/policy-intelligence/docs/iam-simulator-overview - pre-change positioning, 90-day replay, 5,000-log limit, unsupported resource types, unknown results, credential access boundary window hole. Reached via 301 redirect from cloud.google.com; quotes are a fetch-model extraction, not a full-text read.
- https://learn.microsoft.com/en-us/entra/identity/conditional-access/workload-identity - report-only for workload identity policies, single-tenant service principal scope, managed identity and multitenant exclusions, block-only grant, licensing, service principal sign-in logs. Full page text read.
Web, search-summary level only, NOT fetched:
- https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-create-unused.html - unused access analyzer, 1 to 365 day tracking period, service-linked roles not analyzed (an SDK reference page says 1 to 180; unresolved)
- https://docs.aws.amazon.com/wellarchitected/latest/digital-sovereignty-lens/dssec03-bp01.html - rarely used permissions can appear unused in a short window
- https://docs.cloud.google.com/policy-intelligence/docs/role-recommendations-overview - 90-day maximum observation, 30/60-day option, service account worked example
- https://docs.cloud.google.com/policy-intelligence/docs/service-account-insights - service accounts not authenticated in 90 days
- https://docs.cloud.google.com/policy-intelligence/docs/deny-simulator-overview - deny-policy simulator, 90-day replay
- https://docs.cloud.google.com/sdk/gcloud/reference/iam/simulator/replay-recent-access - 1,000-log figure that conflicts with the overview's 5,000
Vault:
~/rdco-vault/06-reference/research/2026-11-06-correct-rule-wrong-default-permission-detection.md~/rdco-vault/06-reference/research/2026-10-24-coded-vs-prompted-agent-permission-gates.md(via parent brief)~/rdco-vault/06-reference/research/2026-10-23-rdco-auto-mode-classifier-miss-profile.md(via parent brief)~/rdco-vault/06-reference/research/2026-07-16-fat-cat-board-router-token-scope.md(title and search snippet only)~/rdco-vault/06-reference/concepts/2026-07-25-agent-first-class-identity.md(search snippet only)