Between 14 and 26 October, a threat actor used IAM credentials from an AWS account associated with Qubole to assume roles that customers had built to trust that account, according to a Wiz write-up produced with Sentra and Mandiant. Every request carried a valid External ID and every one succeeded.
I read that detail as the one that should change how AWS teams write trust policies. Wiz reports that in most of the environments it observed, the actor’s requests were the only AssumeRole activity from the vendor account in October. Most of the roles were not otherwise in use, and the trust policies were still in place. An unused vendor role keeps working for anyone who holds the vendor’s credentials.
Wiz says it has not established how the actor obtained access to the vendor account’s credentials, and I draw no conclusion about the vendor from that. My argument concerns the customer side, where the fix is in our hands. Once inside, according to Wiz, the actor listed Lambda configuration, Secrets Manager and Parameter Store, read from S3 buckets the roles could reach, and tested discovered AWS, Entra ID and Okta credentials. Wiz’s guidance gives the rule I would adopt: add an aws:SourceIp or aws:SourceVpce condition that limits role assumption to the vendor’s documented network ranges. Its write-up says: “In this activity, a source IP condition alone would have blocked every role assumption we observed.”
So my position is that every cross-account vendor role should carry a network condition, and a vendor that cannot tell you its egress ranges should be treated as a finding in the procurement review. The External ID is a weak gate. Wiz notes it is logged in plaintext in CloudTrail and often stored by the vendor and in infrastructure-as-code.
The strongest objection is operational. Vendors change their address space, some publish no ranges at all, and a hard network condition can break a data pipeline at 2 a.m. That cost is real, and it is a plausible reason teams leave the principal as the vendor’s account root. My answer is that the breakage is loud and fixable, while the failure Wiz describes was silent. Wiz found that the actor’s requests came from Mullvad VPN and OVH addresses while all legitimate platform traffic it saw came from AWS infrastructure, which shows the network origin separated the two cleanly. A pipeline that fails closed gets a ticket. A role that works for anyone holding a valid account gets nothing until someone reads CloudTrail.
The same logic argues for deleting roles for integrations nobody uses. Wiz found many victim environments with no legitimate vendor traffic during the activity, and those roles still accepted the actor’s requests. Narrow the principal to the specific IAM principal the vendor uses where the vendor supports it, limit what the role can read, and alert on a vendor account assuming a role from a VPN or hosting provider it has never used.
Wiz also describes a second set of activity involving Qubole integrations, which it has not linked to the role assumptions above. From 8 to 11 October, an actor used customer-owned AWS access keys that Wiz believes were created for Qubole integrations, testing all of them within a single hour on 8 October. In a small number of cases the actor went on to compromise the victim’s cloud environment. Long-lived keys created for an integration are the same problem in a different form: they sit outside any network condition, so they need an owner, an expiry and a review date.
Logging decides how much a victim can learn afterward. Wiz says CloudTrail management events record the role assumption and the bucket listing but not the object reads that make up the data theft, and that in several affected environments missing logs meant responders could not see which S3 objects the actor read. An organization trail covering every account and region, plus S3 data events on buckets that hold sensitive data, turns a guess into an inventory of what left.
This is the same shape we saw when a compromised vendor script drained wallets at Polymarket, and it echoes the case in putting every management interface on a network only administrators can reach: reach is the control, and identity alone is thin. Teams that integrated with Qubole should review CloudTrail from 1 October in every account holding an integration role, using the indicators Wiz published, and rotate anything the role could read if they find a match.
Source: Wiz, “Trusted by Default: Qubole Cross-Account Access Abused in the Wild”