Insights

AI Security

The vendor integration you forgot still has keys to your AWS

In the Qubole campaign, an attacker used credentials from the vendor's AWS account to assume roles customers had created for it, and many of those roles had not been used in months. A one-page inventory of cross-account trust would have shown which doors were still open.

November 18, 2026 · 6 min read

The vendor integration you forgot still has keys to your AWS

A cross-account IAM role keeps working after you stop using the vendor it was built for. Between October 14 and October 26, a threat actor used IAM credentials from an AWS account associated with Qubole to assume roles that customers had created to trust Qubole's account, according to Wiz's write-up, which was done with Sentra and Mandiant. The detail that should bother every security team is this one: in a majority of the environments Wiz observed, the attacker's requests were the only role assumptions from the vendor account that month. The roles were idle. The trust policies were still in place, so the requests succeeded.

SecurityWeek reports that over a dozen organizations had data stolen, and it quotes Sentra: "Qubole roles were still in place long after companies had stopped using Qubole, and those roles could read nearly all of the data the attacker went after."

I am going to stay at the level of what happened and what to do about it. The point for a health system or a HealthTech company is not the mechanics of this campaign. It is that vendor offboarding usually ends at the contract and the invoice, and the cloud permissions keep living.

How the trust was set up, and why it outlived the vendor

Many SaaS platforms that work inside a customer's AWS account ask the customer to create a role during onboarding. That role carries a trust policy allowing the vendor's AWS account to assume it, usually gated by an External ID. When the vendor needs to act, its account assumes the role and gets temporary credentials. This is a normal, documented pattern, and Wiz describes it plainly in its background section.

The trouble is what the pattern assumes. It assumes the vendor's own credentials stay safe, and it assumes someone on the customer side removes the role when the relationship ends. Wiz reports that every role assumption it attributed to the attacker included an External ID, and every one succeeded. Wiz's advice is to treat the External ID as non-secret, because it is logged in plaintext in CloudTrail and often stored by the vendor and in infrastructure-as-code. It was never meant to be the only control.

Then there is reach. According to Wiz, once inside, the actor collected configuration and secrets and read data from S3, opportunistically, "in some cases downloading entire buckets in a single session." Whatever the role could read, the attacker could read.

What would have stopped it

Wiz's recommendations are specific, and one of them is unusually strong: "In this activity, a source IP condition alone would have blocked every role assumption we observed." Adding an aws:SourceIp or aws:SourceVpce condition that limits role assumption to the vendor's documented network ranges means a valid External ID from an unexpected network gets rejected.

The rest of the list reads like a cleanup plan:

  • Inventory every IAM role whose trust policy allows an external account.
  • Delete integration roles you no longer use, or remove the vendor account from the trust policy.
  • Replace a vendor's account-wide :root principal with the specific principal the vendor uses, where the vendor supports it.
  • Cut what integration roles can reach. Wiz notes that access to Lambda configuration, Secrets Manager, SSM parameters and broad S3 read is often granted for convenience.
  • Alert when a vendor role is assumed from a network the vendor has never used.
  • Make sure logging covers every account, and turn on S3 data events or server access logs for sensitive buckets.

That last item matters more than it looks. Wiz points out that CloudTrail management events record the role assumption and the bucket listing but not the object reads, and that in several affected environments, missing logs meant responders could not see which objects were taken. If you cannot answer "what left," your breach notice gets harder to write.

The one-page cross-account trust inventory

Most of the fixes above start with the same question: which outside accounts can act inside ours? I would put the answer on one page, one row per role, and keep it next to the vendor list.

| Field | What to record | | --- | --- | | Role and account | The role name and which of your AWS accounts it lives in | | Vendor | The company the role was built for, and the product | | Trusted principal | Account-wide or a specific principal, and whether a source network condition exists | | Reach | What the role can read or change, in plain words (for example, "read all S3 in prod") | | Last used | When the role was last assumed | | Business owner | A named person who can say whether the integration is still needed | | Contract status | Active, expired, or unknown | | Next review | A date |

The "last used" column is easier to fill than people expect. IAM keeps last accessed information for roles, with a tracking period of at least 400 days for service information. AWS is clear that this data is a guide and that CloudTrail is the authoritative record, so treat a recent date as a reason to check, not proof of compromise.

The column I care most about is "business owner." A role with no owner is a role nobody will delete. If nobody can say why it exists, that is the answer.

Where health systems and HealthTech vendors fit

Health systems extend this kind of trust to HealthTech vendors all the time: analytics platforms, data pipelines, AI features that process records inside the customer's cloud. The Qubole case is a clean argument for asking about it at intake and again at offboarding.

If you are the health system, ask each vendor with a role in your cloud three things. Which principal in your account assumes our role, and from which networks? Will you publish those ranges so we can pin our trust policy? And when our contract ends, who on your side confirms the role is gone from ours?

If you are the vendor, Wiz's last recommendation is written for you. Your principals are "effectively credentials to every customer environment at once." Assume customer roles from workload roles with short-lived credentials rather than long-lived IAM user keys, restrict where those principals can be used, publish your egress ranges, and watch your own logs for role assumptions into customer accounts from unexpected networks. Expect a health system's security reviewer to ask, and have the answer in writing before they do.

What to do this week

Pull the list of roles that trust an external account. For each one, find the owner and the last-used date. Delete the roles for vendors you no longer use. For the ones you keep, add a source network condition if the vendor publishes its ranges, and cut the permissions down to what the integration actually touches.

Then add one line to your vendor offboarding checklist: "Cloud trust removed, confirmed by a named person, with the date." The contract end date is not the day the access ends. Somebody has to make it so.