Document minimum IAM permissions required for self-hosting and BYOC

[waiho-gumloop]

Document minimum IAM permissions required for self-hosting and BYOC

Context

I've been self-hosting E2B on GCP following the self-host guide and needed to determine the exact IAM permissions required for the deployer service account. The guide doesn't document this, and using broad roles like roles/editor isn't an option in our environment due to security policy.

This applies equally to BYOC deployments where the customer's security team needs to approve the IAM role before granting access to their cloud account. Without a documented permission list, the approval process stalls or defaults to overly broad roles.

I ran the full lifecycle (make init → make build-and-upload → make copy-public-builds → make plan → make apply) with a custom IAM role, iteratively adding permissions as 403 errors surfaced and verifying each against GCP audit logs.

Request

Would it be possible to document the minimum required IAM permissions for:

1. GCP (self-hosting and BYOC — one-time init + ongoing CI/CD)
2. AWS (self-hosting and BYOC — same split)

This would help:

• Self-hosters deploying E2B in environments where least-privilege IAM is required
• BYOC customers whose security teams need to approve permissions before granting E2B access to their cloud account
• Enterprise / SOC2 / FedRAMP environments where roles/editor or AdministratorAccess are not allowed

A documented split between one-time init permissions and ongoing CI/CD permissions would also allow customers to grant broader access temporarily during setup, then revoke the init-only permissions afterward.

GCP Permissions (validated end-to-end)

I validated 152 GCP permissions by running the full lifecycle with a dedicated service account and no other roles. These split into two groups:

Init-only (32 permissions) — one-time make init

These are needed only for initial setup (VPC, secrets, buckets, Packer image, Artifact Registry). Not needed for ongoing CI/CD.

Category	Permission	Purpose
IAM	iam.serviceAccounts.create	Create runtime SA for Nomad nodes
IAM	iam.serviceAccountKeys.create	JSON key for runtime SA
IAM	iam.serviceAccounts.getIamPolicy	Read SA IAM to set bindings
IAM	iam.serviceAccounts.setIamPolicy	Grant runtime SA access to buckets/AR
Compute	compute.networks.create	E2B VPC for Nomad cluster
Compute	compute.networks.delete	Packer temp VPC cleanup
Compute	compute.networks.updatePolicy	Add subnets/firewalls to VPC
Compute	compute.subnetworks.create	Subnets within E2B VPC
Compute	compute.subnetworks.delete	Packer temp subnet cleanup
Compute	compute.subnetworks.useExternalIp	Packer VM needs external IP for IAP SSH
Compute	compute.images.create	Packer creates Nomad node disk image
Compute	compute.images.deprecate	Packer sets deprecation status post-build
Compute	compute.disks.useReadOnly	Packer creates image from disk
Compute	compute.disks.delete	Packer temp disk cleanup
Compute	compute.firewalls.delete	Packer temp firewall cleanup
Secrets	secretmanager.secrets.create	~20 secret containers (consul/nomad tokens, etc.)
Secrets	secretmanager.versions.add	Write placeholder values for tokens
Secrets	secretmanager.versions.enable	Enable secret versions after creation
Storage	storage.buckets.create	GCS buckets for templates, kernels, snapshots
Storage	storage.buckets.update	Set versioning and lifecycle rules
Storage	storage.buckets.getIamPolicy	Read bucket IAM for runtime SA binding
Storage	storage.buckets.setIamPolicy	Grant runtime SA objectUser/objectViewer
Storage	storage.hmacKeys.create	HMAC key for ClickHouse GCS access
Storage	storage.hmacKeys.delete	Destroy old HMAC key on recreate
Storage	storage.hmacKeys.get	Terraform refresh of HMAC key state
Storage	storage.hmacKeys.update	Deactivate old HMAC key on recreate
Registry	artifactregistry.repositories.create	AR repos for Docker images
Registry	artifactregistry.repositories.getIamPolicy	Read AR IAM for runtime SA binding
Registry	artifactregistry.repositories.setIamPolicy	Grant runtime SA reader on AR repos
Services	serviceusage.services.enable	Enable GCP APIs (compute, secretmanager, etc.)
IAM	resourcemanager.projects.setIamPolicy	Grant runtime SA project-level roles
IAP	iap.tunnelInstances.accessViaIAP	Packer SSH tunnel to build VM

Ongoing (120 permissions) — CI/CD for make build-and-upload + make plan + make apply

Full list (click to expand)

Category	Permission	Purpose
Registry	artifactregistry.dockerimages.list	Resolve image digests during plan
Registry	artifactregistry.repositories.downloadArtifacts	Pull images onto VMs during apply
Registry	artifactregistry.repositories.get	Refresh repo state during plan
Registry	artifactregistry.repositories.uploadArtifacts	Push Docker images during build
Cert Mgr	certificatemanager.certmapentries.create	Bind cert to domain
Cert Mgr	certificatemanager.certmapentries.get	Read entry state
Cert Mgr	certificatemanager.certmapentries.update	Update entry after cert change
Cert Mgr	certificatemanager.certmaps.create	Create cert map
Cert Mgr	certificatemanager.certmaps.get	Read cert map
Cert Mgr	certificatemanager.certmaps.use	Proxy references map
Cert Mgr	certificatemanager.certs.create	Create TLS cert
Cert Mgr	certificatemanager.certs.get	Read cert provisioning state
Cert Mgr	certificatemanager.certs.update	Update cert after drift
Cert Mgr	certificatemanager.certs.use	Map references cert
Cert Mgr	certificatemanager.dnsauthorizations.create	DNS auth for cert validation
Cert Mgr	certificatemanager.dnsauthorizations.get	Get CNAME for DNS setup
Cert Mgr	certificatemanager.dnsauthorizations.use	Cert references DNS auth
Cert Mgr	certificatemanager.operations.get	Check cert provisioning progress
Compute	compute.addresses.get	Read regional address
Compute	compute.addresses.use	Bind to NAT/LB
Compute	compute.autoscalers.create	MIG autoscaler for node pools
Compute	compute.autoscalers.get	Read autoscaler state
Compute	compute.backendServices.create	LB backend service
Compute	compute.backendServices.delete	Replace backend on update
Compute	compute.backendServices.get	Read backend state
Compute	compute.backendServices.list	Enumerate backends
Compute	compute.backendServices.setSecurityPolicy	Attach Cloud Armor policy
Compute	compute.backendServices.use	URL map references backend
Compute	compute.disks.create	Boot disks for VMs
Compute	compute.disks.get	Read disk state
Compute	compute.disks.list	Enumerate disks
Compute	compute.disks.setLabels	Tag disks
Compute	compute.disks.use	Attach disk to VM
Compute	compute.firewalls.create	Firewall rules for Nomad/Consul
Compute	compute.firewalls.get	Read firewall state
Compute	compute.globalAddresses.create	Global IP for LB
Compute	compute.globalAddresses.get	Read address state
Compute	compute.globalAddresses.list	Enumerate addresses
Compute	compute.globalAddresses.setLabels	Tag addresses
Compute	compute.globalAddresses.use	Forwarding rule references address
Compute	compute.globalForwardingRules.create	LB forwarding rule
Compute	compute.globalForwardingRules.get	Read forwarding rule state
Compute	compute.globalForwardingRules.setLabels	Tag forwarding rule
Compute	compute.globalOperations.get	Poll status of global compute operations
Compute	compute.healthChecks.create	LB health check
Compute	compute.healthChecks.get	Read health check
Compute	compute.healthChecks.use	MIG references health check
Compute	compute.healthChecks.useReadOnly	Backend references health check
Compute	compute.images.get	Read image state
Compute	compute.images.getFromFamily	Resolve image family to latest
Compute	compute.images.useReadOnly	Instance template references image
Compute	compute.instanceGroupManagers.create	Create MIG per node pool
Compute	compute.instanceGroupManagers.get	Read MIG state
Compute	compute.instanceGroupManagers.list	Enumerate MIGs
Compute	compute.instanceGroupManagers.update	Rolling update MIG
Compute	compute.instanceGroupManagers.use	Autoscaler references MIG
Compute	compute.instanceGroups.create	Unmanaged group for backend
Compute	compute.instanceGroups.get	Read group state
Compute	compute.instanceGroups.list	Enumerate groups
Compute	compute.instanceGroups.use	Backend references group
Compute	compute.instanceTemplates.create	VM template per node pool
Compute	compute.instanceTemplates.get	Read template state
Compute	compute.instanceTemplates.list	Enumerate templates
Compute	compute.instanceTemplates.useReadOnly	MIG references template
Compute	compute.instances.create	Create VMs via MIG
Compute	compute.instances.delete	Terminate VMs
Compute	compute.instances.get	Read VM state
Compute	compute.instances.list	Enumerate VMs
Compute	compute.instances.setLabels	Tag VMs
Compute	compute.instances.setMetadata	Startup scripts, SSH keys
Compute	compute.instances.setServiceAccount	Attach runtime SA to VM
Compute	compute.instances.setTags	Network tags for firewall rules
Compute	compute.instances.start	Start stopped VMs
Compute	compute.instances.stop	Stop VMs for maintenance
Compute	compute.instances.use	Attach disks/network
Compute	compute.machineTypes.get	Resolve machine type specs
Compute	compute.networks.get	Read VPC state
Compute	compute.networks.use	Place VMs in VPC
Compute	compute.projects.get	Read project compute metadata
Compute	compute.regions.list	Enumerate regions
Compute	compute.routers.get	Read Cloud Router state
Compute	compute.routers.list	Enumerate routers
Compute	compute.securityPolicies.create	Cloud Armor policy
Compute	compute.securityPolicies.get	Read policy state
Compute	compute.securityPolicies.list	Enumerate policies
Compute	compute.securityPolicies.update	Update Cloud Armor rules
Compute	compute.securityPolicies.use	Backend references policy
Compute	compute.sslPolicies.create	TLS policy (min version, ciphers)
Compute	compute.sslPolicies.get	Read TLS policy
Compute	compute.sslPolicies.use	HTTPS proxy references policy
Compute	compute.subnetworks.get	Read subnet state
Compute	compute.subnetworks.use	Place VMs in subnet
Compute	compute.targetHttpsProxies.create	HTTPS proxy for LB
Compute	compute.targetHttpsProxies.get	Read proxy state
Compute	compute.targetHttpsProxies.use	Forwarding rule references proxy
Compute	compute.urlMaps.create	URL map for LB routing
Compute	compute.urlMaps.get	Read URL map state
Compute	compute.urlMaps.use	Proxy references URL map
Compute	compute.zoneOperations.get	Poll status of zonal compute operations
Compute	compute.zones.get	Resolve zone details
Compute	compute.zones.list	Enumerate zones
IAM	iam.serviceAccountKeys.get	Refresh SA key state
IAM	iam.serviceAccounts.actAs	Attach runtime SA to VMs
IAM	iam.serviceAccounts.get	Refresh SA state
Monitoring	monitoring.metricDescriptors.get	Terraform refresh
Monitoring	monitoring.timeSeries.create	Node SA writes metrics
IAM	resourcemanager.projects.get	Read project info
IAM	resourcemanager.projects.getIamPolicy	Read project IAM bindings
Secrets	secretmanager.secrets.get	Read secret metadata
Secrets	secretmanager.secrets.list	Enumerate secrets
Secrets	secretmanager.versions.access	Read secret values for Nomad jobs
Secrets	secretmanager.versions.get	Read version metadata
Secrets	secretmanager.versions.list	Enumerate versions
Services	serviceusage.services.get	Check APIs are enabled
Services	serviceusage.services.list	Enumerate enabled APIs
Storage	storage.buckets.get	Read bucket state
Storage	storage.objects.create	Upload binaries to GCS
Storage	storage.objects.delete	Replace old binaries
Storage	storage.objects.get	Read objects
Storage	storage.objects.list	List objects / TF workspaces

AWS Permissions (from CloudFormation, unvalidated)

The aws-samples/sample-e2b-on-aws CloudFormation template defines an E2BDeploymentPolicy with ~151 explicit permissions + 5 wildcard groups (autoscaling:*, elasticloadbalancing:*, etc.). These haven't been validated through audit logs like the GCP list.

It would be useful to know if the official iac/provider-aws/ in this repo requires the same permissions as the CloudFormation template, or if there are differences.

Additional observations

• Operation polling permissions: compute.globalOperations.get and compute.zoneOperations.get are needed for Terraform and Packer to poll async operation status. Without them, operations succeed at the GCP API level but the tooling can't confirm completion and reports misleading timeout errors (e.g., Packer's "time out while waiting for image to register").

• compute.images.deprecate is called by Packer after image creation even when no deprecation config is set.

• The init/ongoing split allows BYOC customers to grant broader access temporarily during setup (make init), then revoke the init-only permissions and leave only the CI/CD role for ongoing operations.

Environment

• E2B infra: main branch (commit at time of testing)
• Terraform: 1.5.7
• Packer plugin: hashicorp/googlecompute v1.2.5
• GCP region: us-west1
• Provider: GCP (iac/provider-gcp/)

[linear-code]

ENG-4068