CyberKit

IR playbooks

Stolen Cloud Credentials

A key or a session is being used in your cloud account. Find what it touched and what it left behind.

Cloud credential compromise moves faster than host intrusion because there is no malware to deploy and no lateral movement to perform. An attacker with valid keys is doing legitimate-looking API calls from the first minute, which is why the control-plane log is both the detection surface and the entire investigation.

The first hour

  1. Identify the credential and disable it, rather than deleting it. A deleted identity makes the historical log harder to interpret.
  2. Revoke active sessions. Disabling a key does not end a session already assumed from it.
  3. Pull the full control-plane log for the identity, including the region where global services record, which for AWS is us-east-1.
  4. Look for persistence before looking for damage: new keys, new users, changed trust policies, new service principal credentials.
  5. Check whether logging itself was altered. A trail stopped or its selectors changed marks where visibility ended.

Decisions you will be asked to make

Disable the identity, or watch it?

Options
OptionWhat it costs
Disable nowStops the activity. Tips off the operator, who may already hold other footholds.
Monitor brieflyMay reveal the full extent of their access. Every minute is more damage and more risk.

Disable, unless you have strong reason to believe there are other footholds you cannot yet see and the activity is currently harmless. Monitoring an active intruder is a decision to be made deliberately and with authority, not by default.

How far did the access reach?

Options
OptionWhat it costs
Trust the identity's permissionsFast. Wrong if they escalated, which is the normal case.
Trace every role assumedSlower and accurate. The session chain is what actually bounds the blast radius.

Follow the role chain. In AWS the assumed-role ID stitches the pre- and post-assumption halves of the timeline together; a search for the original user stops at the moment they pivoted.

Was data actually read?

Options
OptionWhat it costs
Data events are enabledYou can answer it directly.
They are notThe log cannot answer it, and the report must say so rather than reporting no access.

Confirm the trail's configuration before drawing any conclusion from absence. Management events are on by default; object-level reads are not.

Scoping questions

  • Which credential was used, and where did it leak from?
  • Which roles were assumed, and what did each of those sessions do?
  • What persistence was created: keys, users, trust policies, service principal secrets, functions, scheduled rules?
  • Was any data made public or shared to an external account?
  • Was logging stopped, narrowed or deleted, and when?

Collect, beyond the standard triage set

  • The full control-plane log for the window, across every region including the one that records global services.
  • The identity's permission history, not just its current permissions.
  • Any resource policies changed during the window, read in full rather than by event name.
  • Snapshot and storage sharing changes, which are exfiltration without network egress.
  • The trail configuration itself, so you can state what would and would not have been recorded.

Known mistakes

  • Pulling only the region where the workload runs, and missing every IAM and STS event because those record centrally.
  • Reading an event name without its parameters. A policy version, a trust policy and a bucket policy all look unremarkable until you read the document.
  • Rotating the compromised key and stopping. Any key, user or trust relationship created during the window is still live.
  • Treating a gap in the log as absence of activity rather than as evidence that logging was interfered with.

On this site