Skip to main content
green gradient background, "The Future of Application Security Is Already Here." and a read the report button.
Qubole's Trusted Roles Became an Attacker's Front DoorPrivileged Access
5 min readFor IAM Architects

Qubole's Trusted Roles Became an Attacker's Front Door

Between October 14 and October 26, an attacker exploited IAM credentials from Qubole's AWS account to assume cross-account roles in customer environments, affecting dozens of customers. This incident raises important questions about the vendor roles in your accounts right now.


Q1: The attacker had the External ID and it still worked. Is External ID useless?

No, but it's not the security control many assume it is. External ID is meant to prevent the confused deputy problem, stopping a vendor's customer A from tricking the vendor into acting against customer B's account. It's not a secret that defends against a compromised vendor. Every malicious role-assumption request in this campaign included the correct External ID and succeeded.

Think of it as a routing check, not authentication. If the vendor's credentials are stolen, the External ID values are available to whoever holds them. Your trust policy must assume the vendor account itself can be compromised.


Q2: What should the trust policy actually look like?

Start with the principal. A trust policy that names an entire external AWS account root allows any identity in that account with sts:AssumeRole permission to access your resources. Narrow trust to specific vendor IAM principals instead. Specify the exact role ARN the vendor's platform uses, not the account.

Then add conditions:

  • Source IP restriction. Use aws:SourceIp to limit assumptions to the vendor's documented network ranges. In this campaign, requests came from a Mullvad VPN address and an OVH-hosted VPS, neither of which is vendor infrastructure. A source IP condition would have rejected them.
  • VPC endpoint condition. If the vendor connects through a VPC endpoint, use aws:SourceVpce.
  • External ID. Keep it. It still blocks confused deputy abuse.

Be aware of one limit. aws:SourceIp doesn't apply to requests made through AWS service endpoints on a caller's behalf, so confirm how your vendor calls the API before you implement the condition.


Q3: We stopped using the vendor months ago. Why does that matter?

Because the role didn't stop existing when the contract did. Many targeted roles had seen no legitimate Qubole activity during October, yet their permissions remained active. The attacker exploited integrations that stayed trusted after some organizations stopped using the platform.

This is Entitlement Drift applied to non-human, cross-boundary access. Unused integrations often go unreviewed, so they keep working. Treat vendor offboarding as a deprovisioning event: delete the role, delete the policy attachments, and remove the External ID from your records. If you can't identify who approved a vendor role and when it was last used, consider it for deletion.


Q4: How do I find every third-party role we've got? We don't have a clean inventory.

Most teams don't. Build one from three angles:

  1. Trust policy scan. Enumerate every IAM role in every account and flag any trust policy whose principal is outside your AWS Organization. IAM Access Analyzer can help with this.
  2. Last-accessed data. For each external-trust role, check when it was last used. Anything dormant is your first candidate for deletion.
  3. CloudTrail AssumeRole history. Group AssumeRole events by calling account and source IP. You'll find vendors nobody remembered approving.

Put the output into your Entitlement Catalog with a named business owner per integration. Then run a Certification Campaign on that list, just as you would for human privileged access. An owner who can't justify a role in writing loses it.


Q5: What did the attacker do once inside, and what should we have been alerting on?

The attack sequence is a useful detection template. The attacker enumerated S3 buckets, databases, EC2 instances, security groups, and IAM identities. They read AWS Secrets Manager, Systems Manager Parameter Store, and Lambda function configurations to pull credentials. They downloaded S3 contents, sometimes entire buckets in single sessions. Harvested credentials, including AWS access keys, Microsoft Entra ID service principal credentials, and Okta OAuth client credentials, were tested with TruffleHog.

Sentra reported over 100 passwords and keys read within minutes. That speed means daily log reviews aren't enough. Build alerts on:

  • AssumeRole from a vendor principal outside the vendor's expected IP ranges
  • Bursts of GetSecretValue and GetParameter calls from a single assumed-role session
  • Bulk GetObject activity across many buckets in one session
  • A vendor role accessing Secrets Manager at all, if the integration has no business reason to

Note that the attacker performed no write operations in the observed environments. If your detection logic focuses on changes, this campaign would have been invisible. Read-only theft is still theft.


Q6: The attacker found Entra ID and Okta credentials in our cloud. How far does the blast radius go?

Beyond AWS. Secrets stored in a cloud account can bridge to your identity provider and SaaS estate. If a vendor role can read Secrets Manager and Parameter Store, then the vendor's compromise is also a compromise of every credential stored there.

Two actions can reduce this risk. First, limit the vendor role's permissions to the specific resources the integration needs, and deny access to secret stores unless necessary. Second, treat any secret a compromised role could read as exposed, and rotate it. This means rotating service principal secrets and OAuth client credentials, and reviewing what those identities can access. Prefer short-lived, federated workload credentials over long-lived keys where possible, so a stolen secret expires on its own.


Where to go for more

If you use Qubole, review AWS CloudTrail records back to October 1, 2026, across every account containing integration roles. Look for suspicious AssumeRole activity originating from Qubole's AWS account, especially from non-AWS infrastructure, then trace follow-on activity in Secrets Manager, Parameter Store, Lambda, and S3.

For a general fix, start with IAM Access Analyzer external access findings and the AWS documentation on the confused deputy problem and cross-account role conditions. Make third-party role review a regular item in your Certification Campaign calendar, not a one-time cleanup.

Promotional banner highlighting failures found in PCI audits and how to spot the gaps

You Might Also Like