All articles

NewsNovember 9, 20265 min read

Qubole: trusted AWS roles abused to steal customer secrets

IAM credentials tied to Qubole’s AWS account were used to assume customer roles: the attack chain, indicators and the CloudTrail checks to run now.

Blue carabiner with the word “Safety” clipped to a red strap

An unidentified attacker used IAM credentials tied to the AWS account of Qubole, a cloud platform for processing and analysing large volumes of data, to get into the AWS accounts of customers that had configured a trusted cross-account IAM role for the vendor. From there they read secrets, tested the credentials they found and downloaded data from Amazon S3. The campaign was flagged by Sentra, which saw suspicious activity in some of its customers’ environments, and analysed by Wiz together with Mandiant. AWS states that its services were not compromised and operated as designed.

What happened

To manage compute and storage in the customer’s account without long-lived access keys, Qubole relies on IAM roles that the customer creates during onboarding: the trust policy lets identities in Qubole’s AWS account assume them with sts:AssumeRole, with an External ID as an additional check. The attacker held credentials that could satisfy that check too: according to Wiz, every malicious AssumeRole request carried the correct External ID and succeeded.

The weak spot was forgotten integrations. Many roles showed no legitimate Qubole activity in October, and several companies had left them in place after they stopped using the platform, with read access to almost all the data the attacker went after.

How Qubole’s credentials were obtained has not been disclosed, and when the sources were published Qubole had not replied to the press or issued an advisory. The activity is not attributed to any known group.

The attack chain

  • Initial access: on 14 October a Mullvad VPN address tries to assume customer integration roles from Qubole’s AWS account 805246085872, using IAM user AIDA3W7DIFLYOLTK63SFS. From 18 October the attacker uses a second IAM user in the same account, alongside another Mullvad address. Most of the later activity comes from an OVH-hosted VPS, 57.129.54[.]120. According to Sentra, none of the requests came from Qubole’s network. The campaign ran from 14 to 26 October.
  • Discovery: enumeration of S3 buckets, databases, EC2 instances, security groups and IAM identities. Role session names follow a fixed pattern, the prefix quiam- followed by the action, for example quiam-enum and quiam-scope-full.
  • Secret harvesting: reads from AWS Secrets Manager, Systems Manager Parameter Store and Lambda function environment variables. Sentra reports more than 100 passwords and keys read within minutes (seven, according to BleepingComputer) and thousands more credentials inside the downloaded files.
  • Credential validation: with the TruffleHog tool the attacker validated AWS keys, Microsoft Entra ID service principal secrets and Okta OAuth client credentials, which were also used to request tokens.
  • Exfiltration: always the same sequence, ListBuckets, ListObjects and HeadObject, then GetObject. In some cases entire buckets were downloaded in a single session; according to Sentra one customer alone lost millions of files. No write operations were observed: the goal was data theft, with an opportunistic rather than targeted approach.

On the attacker’s OVH infrastructure Wiz also found a bootable ISO image, sso_agent.iso, which collects system information, talks to the server over TCP port 23458 and runs shell commands received remotely; its configuration suggests Windows systems as the target. Wiz found no use of it against victims.

Wiz also describes a separate campaign between 8 and 11 October, in which customer-owned AWS access keys used for Qubole integrations were tested against victim environments, successfully in a small number of cases. A link with the main campaign has not been established.

In our reading the relevant MITRE ATT&CK techniques are T1199 (Trusted Relationship), T1078.004 (Cloud Accounts), T1580 (Cloud Infrastructure Discovery), T1555.006 (Cloud Secrets Management Stores), T1552 (Unsecured Credentials) and T1530 (Data from Cloud Storage).

How many victims

The sources differ. Wiz speaks of multiple affected environments without giving a number; SecurityWeek writes of over a dozen organisations, while Sentra, quoted by CSO Online and BleepingComputer, speaks of dozens of Qubole customers. Victim names have not been made public.

Impact for European and Italian companies

Anyone who has used Qubole, even only in the past, should treat the AWS account as potentially exposed for as long as the integration role exists. But the issue is wider: many SaaS services work inside customer accounts through cross-account roles, and an unused role with broad permissions is permanent access that depends entirely on the vendor’s security. For NIS2 entities, supply-chain security is one of the required measures, and unauthorised access through a vendor should be assessed as a possibly reportable incident; if the buckets held personal data, the GDPR breach notification within 72 hours must also be considered. Entra ID and Okta credentials stored in AWS extend the exposure beyond the original cloud.

What to do now

  1. Find the roles that trust Qubole across all AWS accounts by checking trust policies for account 805246085872.
  2. Review CloudTrail from 1 October 2026: AssumeRole events where userIdentity.accountId equals 805246085872, especially from non-AWS addresses, and session names starting with quiam-.
  3. For every session found, trace all subsequent activity, in particular on Secrets Manager, Parameter Store, Lambda and S3.
  4. If the role was assumed, rotate every secret, parameter and Lambda environment variable the role could read, including Entra ID and Okta credentials, and revoke the role’s active sessions: rotation alone does not end temporary credentials already issued.
  5. Remove integrations no longer in use and cut the permissions of the remaining ones to the minimum needed.
  6. Tighten trust policies: name the vendor’s specific IAM principals instead of the whole external account, and restrict role assumption to the vendor’s documented addresses with source IP or VPC endpoint conditions.
  7. Hunt for the indicators 57.129.54[.]120 and TCP port 23458 in network and firewall logs.

How ARXDOME helps

ARX Horizon lets you query threat intelligence sources about the published indicators and see, for each address, country, operator, VPN or cloud use and reputation; the same data enriches the events the platform collects, such as connections logged by firewalls. The CloudTrail and trust policy checks described here still have to be done in the AWS console.

Last updated:

Any questions?

Let’s talk about your company’s needs, with no obligation.

Contact us

Related articles