- AWS IAM
- Azure RBAC
- Google Cloud IAM
One request · One review · One role · Every cloud
Cloud access starts simple. Then it is everywhere — access sprawl.
A handful of accounts. Directory groups mapped to a few cloud roles. One person who knows how it all fits together. It works, because it is small.
Hundreds of accounts, subscriptions and projects. A daily queue of access tickets. Hand-written policies nobody dares to touch. People move on; their access stays. And now the pipelines and AI agents are asking too.
Who decides what a role may contain — and who checks it still only contains that?
How do you give the same answer on AWS, Azure and Google Cloud, when each works differently?
How does a request get reviewed without a week of chasing approvers?
What takes access away when it is no longer used, and how do you prove it happened?
What stops a machine or an agent from quietly collecting standing power?
AI agents never sleep.
Pipelines, service accounts and AI agents already hold more cloud access than most people. They ask at machine speed, nobody sponsors them, and nothing adds up what they hold in total.
Code & configs
Agents can read, change and exploit applications and infrastructure faster than a reviewer can follow.
Standing access
A long-lived key or role becomes permanent attack surface the day after it is created.
Machine speed
A review process built for a ticket a day cannot keep pace with a thousand automated requests.
Good cloud tools exist today — but they all assume the role already exists.
The access market has three mature categories. Each solves a real problem, after the fact: they find, elevate into or guard a role somebody already built. None of them decides what that role should contain, proves it stays bounded, or retires it with evidence.
Show who has what
Entitlement scanners map every permission across your clouds and flag the risky ones. They show the mess. They do not change it.
e.g. Wiz · Microsoft Defender for Cloud · Sonrai · Veza
Switch access on for a while
Just-in-time and privileged-identity tools let someone step into a role for an hour, with approval. The role itself — and whatever is inside it — is taken as given.
e.g. Microsoft Entra PIM · Britive · Apono · ConductorOne
Guard the credentials
Secret vaults and session brokers protect the keys and record the session. They do not ask whether the key should open that door at all.
e.g. CyberArk · HashiCorp Vault · BeyondTrust · Teleport
Makes the role right before anyone is elevated into it
Bastion owns the layer underneath: what a role is allowed to contain, how it is generated, who reviews it, how it reaches each cloud, how it is recertified and how it is retired. It works alongside the tools above — a role Bastion built is the one your elevation tool switches on and your scanner finds clean.
One engine for the whole life of a cloud role.
In plain words: someone asks for access, Bastion writes the smallest role that fits, routes it to the right reviewers, deploys it safely, recertifies it on a schedule, and removes it when it is no longer needed. Every step leaves a record — on AWS, Azure and Google Cloud alike.
AWS IAM, Azure RBAC and Google Cloud IAM work differently under the hood. Bastion builds each role the native way, through the same engine, the same review and the same evidence — so a role means the same thing on every cloud.
Requests start from versioned, reviewed role templates. Wildcards, self-grants and privilege-escalation paths are refused before anyone sees the request.
Risk decides who reviews. Low-risk changes take the fast lane; elevated ones add a manager; high-risk ones add security. Nobody approves their own request.
Every role is recertified on a schedule set by its risk. Expired grants — and on AWS, provably unused ones — step down a decommission ladder, and the evidence that access ended is kept.
- One engine, three clouds
- 0 standing cloud keys — short-lived OIDC credentials only
- Sandbox first, then the target, with checks after each
- Permission boundaries on every AWS role
- Recertification every 30, 90 or 180 days by risk
- Tamper-evident, hash-chained audit trail
- The portal cannot write to a cloud — only the deploy pipeline can
- Controls are code, CI and cloud settings — not process documents
Three clouds that work differently. One role that means the same thing.
AWS, Azure and Google Cloud each model a role their own way, and Bastion builds each one natively — consistently, from one request. Every cloud gets a reviewable artifact and the same four checks after deployment: it exists, it grants what was approved, it grants nothing forbidden, and it cannot escape its scope. Adding a cloud adds a bundle; the engine does not change — Google Cloud and Azure were added without touching the core.
One role with a permission policy, a trust policy and a permissions boundary — as Terraform, under review.
No iam:* and no self-granting IAM writes. No public or loose trust. Boundary attached, always.
Role read back after apply, then linted by the cloud's own policy analyzer.
One custom role definition (both permission planes) and one assignment to one named principal, at subscription scope.
No wildcard operations — Azure allows them, Bastion refuses them. No subtractive rules. No role or policy writes. Never a management group or root.
Definition and assignment read back; live permissions compared against what was approved, on both planes.
One project-scoped custom role and one binding to one named member.
No wildcards. No permissions that change IAM or mint keys. Never allUsers or allAuthenticatedUsers.
Role and binding read back; live permissions compared against what was approved; member confirmed not public.
Where the cloud itself cannot cap a grant, Bastion says so and routes every request through a human reviewer. It never pretends a guardrail exists where one does not.
Five verbs. One path.
Every request walks the same spine, and gets narrower at each step.
Request
A guided form, not a ticket. The policy previews as you type, and every fixable problem is named before you can submit.
Generate
Nothing is hand-written. The role is rendered from a versioned template into reviewable code.
Approve
Risk picks the reviewers. Separation of duties is enforced, and the portal is the only approval authority.
Deploy
Sandbox first, then the target. No standing keys. Safe to re-run.
Prove
Hash-chained, tamper-evident. Every step leaves an artifact an auditor can read.
And after it is live
Recertify
Owners re-attest on a cadence set by risk — 30, 90 or 180 days. On AWS, provably unused roles are flagged from real last-used data.
Retire
Expired grants step down a ladder: soft-disable, a grace period, then removal. The final delete is done by a person, never silently.
Prove it ended
Removal is recorded in the same chain as the grant, so an auditor can follow a role from first request to last breath.
Doors, not keys.
Pipelines, services and AI agents request access through the same engine as people — via a machine API or an MCP endpoint — and walk the same five steps.
Standing keys. Long-lived roles. Broad permissions that accumulate. No velocity limits. Agents inherit the same problems as humans — only faster.
Temporary doors. Every machine grant carries an expiry and a named human sponsor, with velocity and budget caps so no principal can quietly accumulate power. Least privilege by construction for every principal — human or machine.
Overstate nothing
The audit chain is hash-chained and tamper-evident: each entry commits to the one before it, so a deletion or a reorder is detectable. It is not signed and not anchored — those are not switched on, so we do not claim them. We say tamper-evident, never immutable. Claims match mechanisms, including this one — and the register that keeps this page honest is checked on every build.
Where it is going
- Rule-based delivery: approve a rule once, and Bastion keeps thousands of accounts, projects and subscriptions in line with it.
- AWS Identity Center for people: permission sets and groups instead of per-account roles — proven in the lab, not yet switched on.
- Groups born in your directory: Bastion creates the group, your directory decides who is in it.
See it on your own terms.
The safe path can also be the fast path. Tell us about your team and we'll walk you through Bastion on sample data — no cloud account required, no sales runaround.
Prefer email? Reach the platform team directly. info@cygnustechnologies.net