Between October 14 and October 26, an attacker used IAM credentials from Qubole's AWS account to assume roles in customer environments. These roles were set up for legitimate integrations, the External ID check passed, and dozens of customers were affected. Many roles had seen no legitimate Qubole activity during October, yet their permissions remained active.
This checklist is for procurement security professionals managing vendor access to cloud environments. Each item has a clear done/not-done state and describes what effective management looks like. It aligns with Third-Party Risk Management (TPRM) workflows, and you can reference ISO 27036 and SR 23-4 when documenting evidence.
Prerequisites
Before you start, confirm you have:
- Read access to AWS CloudTrail across every account with an integration role.
- An inventory of vendors with any form of delegated cloud access, owned jointly by procurement and security.
- A named owner for each vendor relationship who can confirm whether the integration is still in use.
- The authority to disable or delete roles, or a direct escalation path to someone who has it.
The Checklist
1. Build a complete inventory of cross-account roles that trust external AWS accounts.
Done: You've enumerated every IAM role whose trust policy names an AWS account outside your organization, and each one maps to a named vendor and business owner. Not done: Any role exists with no known owner.
Effective management looks like: The inventory is generated from your cloud configuration, not recalled from memory, and is refreshed on a schedule.
2. Remove integrations for vendors you no longer use.
In this incident, the attacker exploited integrations that remained trusted even after some organizations stopped using the platform. Done: Every role is tied to an active contract or has been deleted. Not done: A role outlives its commercial relationship.
Effective management looks like: Access Revocation is a required step in your Wind-Down Plan, with role deletion confirmed in writing by security.
3. Narrow trust policies to specific vendor principals.
Trust specific vendor IAM principals rather than allowing an entire external AWS account to assume a role. Done: No trust policy names a vendor's account root. Not done: Any policy lets every identity in the vendor's account assume your role.
Effective management looks like: The trust policy lists named roles or users, so a single compromised credential elsewhere in the vendor's account can't satisfy it.
4. Treat the External ID as one control, not the control.
All malicious role-assumption requests in this attack included the required External ID and succeeded. Done: Your design assumes the External ID can be satisfied by an attacker holding vendor-side credentials, and other controls compensate. Not done: The External ID is your only barrier.
Effective management looks like: You've layered source restrictions, session limits, and monitoring on top of it.
5. Restrict role assumption to documented vendor network ranges.
Use AWS source IP or VPC endpoint conditions. Here, activity came from a Mullvad VPN address and an OVH-hosted VPS, neither of which looks like a vendor's production network. Done: Trust policies carry network conditions backed by ranges the vendor has documented in writing. Not done: Assumption works from any IP on the internet.
Effective management looks like: The vendor is contractually obligated to notify you before their ranges change.
6. Apply least privilege to what the role can do.
The attacker enumerated S3 buckets, databases, EC2 instances, security groups, and IAM identities, then read Secrets Manager, Parameter Store, and Lambda configurations. Done: Each role's permissions are scoped to the specific resources the integration needs. Not done: The role can list or read broadly.
Effective management looks like: An access review shows the role can't reach secrets stores or unrelated buckets.
7. Keep secrets out of reach of vendor roles and out of plaintext files.
Credentials found in compromised environments included AWS access keys, Microsoft Entra ID service principal credentials, and Okta OAuth client credentials. Done: Vendor roles can't read credential stores, and a scan of accessible files finds no live secrets. Not done: Downloadable data contains keys.
Effective management looks like: Secrets are rotated on a schedule and immediately after any vendor-side incident.
8. Alert on assumption of vendor roles from unexpected sources.
Done: You have detections on AssumeRole events for integration roles, flagging non-vendor source IPs and dormant roles that suddenly activate. Not done: You only review logs after an advisory.
Effective management looks like: A dormant role being assumed for the first time in weeks pages a human.
9. Review CloudTrail history after a vendor disclosure.
Review CloudTrail records back to October 1 across all accounts containing integration roles, and look at follow-on Secrets Manager, Parameter Store, Lambda, and S3 activity. Done: You can run that lookback across every account within hours. Not done: Logs are retained in only some accounts or for too short a period.
Effective management looks like: Retention covers the lookback window an advisory is likely to demand, centrally and in every account.
10. Contract for Vendor Breach Management.
Done: Agreements define a Notification Timeline, require disclosure of compromised vendor credentials, and include a Right to Audit covering the vendor's access-key handling. Not done: Contracts are silent on credential compromise.
Effective management looks like: You can name the clause that obligated the vendor to tell you, and the date they must do so by.
11. Include vendor access in Continuous Monitoring of Active Arrangements.
Done: Access posture is reviewed on a set cadence alongside other signals such as Cyber Risk Rating changes. Not done: Access is reviewed only at renewal.
Effective management looks like: Criticality Classification drives review frequency, and vendors with read access to production data are reviewed most often.
Common Mistakes
- Treating contract end as access end. Roles persist until someone deletes them.
- Trusting an account instead of a principal. One leaked credential then unlocks every customer.
- Relying on the External ID alone. It protects against confused-deputy problems, not stolen vendor credentials.
- Reviewing access only at renewal. Dormant, over-permissioned roles are what attackers look for.
- Keeping logs in one account only. A lookback is only as good as your retention in the accounts that matter.
Next Steps
Start with items 1 and 2; they cost the least and remove the most exposure. Then prioritize items 3 and 5 for any vendor with data access, since they directly limit what a stolen vendor credential can do. Finally, feed the results into your Risk Scoring so vendors with standing, broad, unmonitored access are visible to the people who approve renewals.





