Compliance

Phase 1, Phase 2, then keep it running

A fixed-price gap analysis, remediation through to your audit, and upkeep after it. All in a workspace you keep.

All compliance →
Security

Testing, review and leadership

Led by a published security researcher with five CVEs. A findings report, a letter of attestation for your buyers, and a retest of your fixes.

All security →
Who we help

Prove you are secure

To the people you sell to, raise from or answer to.

All industries →
Resources

Learn the space

Original research, free tools, and plain-language guides on security and compliance, from a published security researcher.

Read the blog →
Security

How to Audit Vendor Access to Your AWS Account

To audit vendor access to your AWS account, start from your own IAM configuration rather than from your vendor list: enumerate every role whose trust policy names a principal outside your organization, every IAM user created for a third party, and every resource policy on S3, KMS and Secrets Manager that grants an external account. Then, for each one, compare the permissions you granted against the API calls the vendor has actually made in the last ninety days and cut the grant down to that. This is a half-day exercise for most companies with twenty to two hundred staff, and it finds things no vendor questionnaire will ever surface.

The reason this matters right now: Wiz reported that between October 14 and October 26, 2026, an attacker used credentials for an AWS account belonging to the data lake platform Qubole to reach into the AWS accounts of multiple Qubole customers, taking stored secrets and data. How the attacker got those credentials has not been disclosed, and from the customer side it barely matters. The customers were breached through a door they installed themselves.

What vendor access to your AWS account actually is

Most SaaS platforms that work with your cloud infrastructure do it one of three ways. The cleanest is a cross-account IAM role: you create a role in your account whose trust policy allows a principal in the vendor's AWS account to assume it, usually with an external ID to prevent the confused deputy problem. The second is a long-lived IAM user with an access key, which the vendor stores somewhere in their systems. The third is a resource-based policy, where you add the vendor's account to a bucket policy or KMS key policy directly, with no role involved.

All three are standing access into production. A SOC 2 Type II report from the vendor tells you the vendor has a change management process and does background checks. It tells you nothing about whether the role you created for them permits s3:GetObject on every bucket in the account, or whether the trust policy names the vendor's account root instead of a specific role, which means any principal that vendor creates, or any attacker who gets into that vendor's account, can assume it.

That is the gap. The control that fails is not the vendor's, it is yours. Your blast radius was set the day an engineer pasted a CloudFormation template from the vendor's onboarding docs and clicked create stack without reading the policy document inside it.

How to audit vendor access to your AWS account, step by step

Do this in order. Each step is a query or a short script, and the output of each one feeds the next.

  • Enumerate roles with external trust. List every IAM role in every account in the organization and parse the trust policy document. Flag anything where the Principal block names an AWS account ID that is not yours, or where it is a wildcard. Do this per account, not just in the management account.
  • Run IAM Access Analyzer with the zone of trust set to your organization. It will surface external access on roles, S3 buckets, KMS keys, Secrets Manager secrets, SQS queues and Lambda functions. Export the findings. This catches the resource-based grants that a role-only review misses.
  • Find vendor IAM users. List IAM users with active access keys, check key age, and identify which ones belong to a vendor integration rather than a person. Any key older than a year that belongs to a third party is a finding on its own.
  • Attribute each grant to a vendor, an owner and a reason. If nobody internally can name why a role exists, that is your answer. Roles left over from a trial you abandoned eighteen months ago are the most common thing we find.
  • Check the trust conditions. Each role should name a specific role ARN in the vendor's account, not the account root. It should carry a unique external ID that the vendor generated per customer. It should have a session duration that matches the work being done, not the twelve hour maximum.
  • Compare granted to used. Pull the last ninety days of CloudTrail for AssumeRole events on each vendor role, then pull every API call made under those session identities. IAM access advisor service last accessed data gives you a faster first pass at the service level. You now have the real list of what the vendor calls.
  • Cut the policy down. Rewrite the permission policy to the services and actions that appear in the log, scoped to named resource ARNs where you can. Keep a short allowance for actions the vendor uses rarely, but remove entire services that have never been touched.
  • Add detection. Alarm on changes to the trust policy of any vendor role, on AssumeRole volume outside the normal pattern, and on the vendor session calling a service it has never called before. Decide now who revokes access and how, so that during an incident the answer is a documented step rather than a debate.

While you are in CloudTrail, check what the vendor role could reach if it were abused. If the role can read Secrets Manager or SSM Parameter Store, assume everything in there is exposed in a vendor compromise. That is exactly what was reported stolen in the Qubole case.

How to decide what permissions a vendor role should actually have

The vendor's own onboarding template is a starting negotiating position, not a requirement. Vendors write broad policies because broad policies generate fewer support tickets. We have cut vendor roles by eighty percent of their granted actions with no functional impact, and the vendor never noticed.

A few decision rules that hold up:

  • If the integration is read-only in nature, say log ingestion or cost analysis, there should be no write or delete action anywhere in the policy. AWS managed ReadOnlyAccess is not a safe default here, because read on an object store means read of the data itself.
  • If the vendor needs write access to one bucket, the policy names that bucket ARN. No wildcards on resource for data plane actions.
  • Anything touching IAM, KMS key policies, organization settings or CloudTrail configuration needs a written justification from the vendor describing the exact call. Most of the time there isn't one.
  • Separate roles per environment. A single role trusted across production and staging means a vendor compromise is automatically a production compromise.
  • Where the vendor supports it, add a permissions boundary or an SCP that caps what any third-party role can do regardless of its attached policy. This is the control that still works when someone grants too much in six months.

If the vendor refuses to operate within a scoped policy and will not explain why, that is a risk acceptance decision for a named owner, recorded in writing, not a technical problem. The same logic applies to tools that act on your systems rather than just answer questions, which is why assessing a vendor that acts rather than just answers has become its own category of review.

Free weekly email

Get The Compliance Brief every Tuesday

One email a week from Jacob Masse: the security and compliance stories that changed something that week, and what each one means if you sell software to enterprise buyers. Five stories, a take on each, five minutes to read.

Free. Unsubscribe in one click, and replies reach Jacob directly. Read the latest issue or browse the archive.

What evidence an auditor or enterprise buyer will ask to see

Auditors test supplier management under SOC 2 CC9.2 and logical access under the CC6 series. ISO 27001 pushes it through the supplier relationship controls in Annex A clause 5. In both cases the examiner wants to see that third-party access is inventoried, approved, scoped and reviewed, and the artefacts that satisfy that are specific:

  • A register of every third-party access path into your cloud, with vendor name, account ID, role ARN, owner, business purpose, external ID present yes or no, and date last reviewed. This belongs alongside, or inside, the subprocessor inventory you maintain for the audit.
  • Evidence of the review itself: the Access Analyzer export, the CloudTrail query results, and the dated ticket or document showing who reviewed and what changed. A register with no supporting review output reads as a spreadsheet somebody typed once.
  • Approval records showing a named person authorised the grant before it was created, and offboarding records showing roles deleted when vendors were dropped.
  • The monitoring configuration: the alarm definitions for trust policy changes and anomalous assume-role activity, plus one example of an alert firing and being handled.

Run this review on a defined cadence, quarterly for anything touching production, and keep the output filed the same way you file user access reviews. The mechanics are close enough to running an access review that passes a SOC 2 audit that you can reuse the template and the sign-off workflow.

The mistakes that cause findings

The ones we see repeatedly in readiness work:

  • Trusting the vendor's account root. It means every principal in their account, now and in future, can assume your role. Name the specific role ARN.
  • No external ID, or an external ID that is the same for every customer. Ask the vendor how they generate it. If they cannot answer, that is a signal about the rest of their engineering.
  • Treating the vendor's SOC 2 report as the control. Collecting a report and filing it does nothing about the permissions you granted. The report scope almost never includes the way the vendor stores the credentials that reach into you.
  • Forgetting the trial. Proof-of-concept roles outlive the proof of concept. Add role deletion to your vendor offboarding checklist, and verify it, because deleting the SaaS account does not delete the IAM role.
  • No logging on the vendor's session. If you cannot tell an investigator what the vendor role did on a given day, your incident response is guesswork. This is the point at which an incident management process built in advance stops being paperwork.
  • Secrets reachable from the vendor role. Audit what the role can read in Secrets Manager and Parameter Store specifically, and rotate anything it could have read if you ever suspect vendor compromise.

How this differs by company size and framework

At twenty to fifty people with a single AWS account, you may have three or four external grants and the whole review fits in an afternoon with no tooling beyond the CLI. The risk is concentration: one role usually has far more than it needs because there was never a reason to split it.

Between fifty and two hundred people with an organization and multiple accounts, the problem shifts to discovery. Roles get created in accounts nobody is watching, by teams who onboarded a tool without telling security. Here you want Access Analyzer running continuously at the organization level and an SCP capping third-party role permissions, because point-in-time review will not keep up with the rate of change. That slow divergence between what you documented and what is running is the same pattern described in control drift between audits.

On frameworks: SOC 2 gives you latitude to define the control, so write the description to match what you will actually do quarterly, not an aspirational monthly review you will miss. ISO 27001 expects the supplier agreement itself to address access, so the contract and the IAM configuration need to agree. If you are in a regulated sector, the expectations around documenting and monitoring third-party access are stricter still, and third-party risk management for fintech covers where those extra obligations land. Before any of it, it is worth doing a full pass over the account itself, which is the point of pre-auditing your cloud before your auditor does.

Common questions about auditing vendor access to your AWS account

How often should we review cross-account roles?

Quarterly for any role that touches production, annually for the rest, plus an unscheduled review whenever a vendor discloses a security incident or you change your account structure. Tie the quarterly run to the same calendar as your user access review so one person owns both and the evidence lands in the same folder. If you are mid observation window, keep the cadence consistent, because a skipped quarter shows up as a gap in the period.

Does a vendor's SOC 2 report cover the access they have in our account?

Almost never in the way buyers assume. The report covers the vendor's own system and the controls they described in scope, which typically means their application and infrastructure, not the specific credentials and roles they hold in each customer environment. Read the system description and the complementary user entity controls section, because that section is where the vendor tells you which responsibilities are yours, and scoping the role you create is usually one of them.

What is an external ID and why does it matter?

An external ID is a shared secret included as a condition in the role's trust policy, so that a vendor can only assume your role when they present that value. It prevents the confused deputy problem, where an attacker who knows your role ARN convinces the vendor's platform to assume it on their behalf. It should be unique per customer and generated by the vendor, not guessable like your company name, and if a vendor's integration does not offer one, raise it with them in writing and record the decision.

What should we do if a vendor with AWS access reports a breach?

Revoke first, investigate second: attach a deny-all policy to the role or remove the trust relationship, then pull CloudTrail for every session under that role across the disclosed window and look for calls outside the normal pattern. Rotate every secret the role could read, including anything in Secrets Manager and Parameter Store within its resource scope, and treat data in buckets it could read as potentially accessed until the logs prove otherwise. Keep the timeline written down as you go, because customer notification obligations and your own liability position as a SaaS vendor in a customer data breach will both depend on what you can evidence about timing and scope.

This started as a story in The Compliance Brief. Every Tuesday, Jacob picks the five security and compliance stories that change something for a company selling to enterprise buyers, and says what to do about each. This one was in issue 15.

Get the Brief every Tuesday

Before you go

Want the rest of this by email?

If this was useful, I send a few short notes on vendor risk and security questionnaires. Unsubscribe in one click, and replies reach me directly.

From Jacob Masse, principal of traztech: the files by email, then a few short notes over the next month. No spam, unsubscribe in one click.

Want a second opinion on where you stand?

We run SOC 2, ISO 27001 and the rest of the compliance stack for startups and SMEs, and the security testing that sits behind it. The first call is free, and we will tell you if you are not ready to start yet.

Book a free call