Skip to main content
Back
IntrusionAWS · SaaS
10 min read

AWS intrusion and extortion investigated by Sygnia and described as AI-assisted

Victim
An unnamed organization with a large AWS environment
When
Not stated; disclosed June 2026Disclosed Jun 29, 2026
Entry
An AWS access key obtained through an unnamed weakness in an internet-facing application.

In an intrusion Sygnia investigated and described as an AI-assisted cloud attack, an actor obtained an access key to one AWS account through a weakness in an internet-facing application. Each new key was checked, reused in its own account and used to find more: secrets in Secrets Manager, Parameter Store, S3, workload environment variables and GitHub and Bitbucket runners. The actor created IAM users and keys, opened reverse shells, backdoored the delivery chain, queried dozens of RDS databases and used mostly reversible outages as extortion leverage; Sygnia calls the evidence of AI assistance consistent with it but not definitive.

What happened

Internet
AWS APIstolen access keys
Application and Workloads
Data and Secret Stores
+ runner environments
+ deployment files
+ images and webhooks
GitHub · Bitbucket
Source Control and CI/CD
Internet-Facing Application
EC2 and ECS Workloads
Secrets Manager, Parameter Store
New IAM Users
S3 Buckets
SQS Queues
RDS Databases

Replaying the Sygnia June 2026 intrusion

  1. 1

    A key from the application

    Day not stated

    A weakness in an internet-facing application gives up an access key to one AWS account, and every new key is checked and reused in its own account.

  2. 2

    Secrets, then more keys

    Within about 72 hours of entry

    Stolen keys read Secrets Manager secrets, Parameter Store parameters and plaintext secrets in S3; workload environment variables and CI/CD runners give up more.

  3. 3

    Footholds of its own

    Within about 72 hours of entry

    The actor creates IAM users and access keys, opens reverse shells on EC2 instances and ECS containers, and backdoors the delivery chain.

  4. 4

    Customer data

    Within about 72 hours of entry

    Several hundred unique SQL queries run across dozens of RDS databases for user and transaction data; how the actor connected is not stated.

  5. 5

    Outages as leverage

    Within about 72 hours of entry

    ECS services go to zero capacity, SQS queues are purged, S3 buckets are denied to their owners and network access is blocked, all mostly reversible.

Impact

Broad compromise of a large AWS environment in about 72 hours: secrets read across it, dozens of RDS databases queried for user and transaction data, and mostly reversible outages of S3, ECS, SQS and network access used as extortion leverage.

Outcome

Where cloud-provider controls fit

No cloud-provider control reaches a weakness that no source names, so the first key stays the application's problem. Once a key is out, two controls would have stopped a step outright. A service control policy with a network condition refuses every call that a stolen key signs from outside your own networks, apart from the identity check, if the accounts are member accounts of your organization; the write-up does not say where the calls came from. A service control policy that keeps user, access key and password creation to the identity administrators' role refuses the new IAM users and keys, if the stolen key is not that role's. The other controls bound a step. Resource control policies would have kept Secrets Manager and S3 reads inside your networks, but they do not reach Parameter Store or the secrets held in workloads and runners. Service control policies that keep capacity changes, queue purges and bucket and network rule changes to named roles stay caps, because the write-up does not name the calls the actor used. Security groups would have bounded the reverse shells and the database connections, and a GitHub ruleset would have put a reviewer in front of a backdoored commit. Secrets kept in plaintext in workloads and runners stay readable to whoever reaches them.

Could this have been prevented?

In hindsight, not as a whole. 2 controls would have refused a step outright, and 9 controls would have limited one. 3 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.

2
Prevented
9
Contained
3
Not covered

How it could have been prevented

The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.

Internet
AWS APIstolen access keys
Application and Workloads
Data and Secret Stores
+ runner environments
+ deployment files
+ images and webhooks
GitHub · Bitbucket
Source Control and CI/CD
Internet-Facing Application
EC2 and ECS Workloads
Secrets Manager, Parameter Store
New IAM Users
S3 Buckets
SQS Queues
RDS Databases

Replaying the Sygnia June 2026 intrusion

1 · A key from the application

1 Prevented1 Contained1 Not covered6 on other clouds

2 · Secrets, then more keys

2 Contained1 Not covered6 on other clouds

3 · Footholds of its own

1 Prevented2 Contained1 Not covered6 on other clouds

4 · Customer data

1 Contained3 on other clouds

5 · Outages as leverage

3 Contained6 on other clouds

Key facts

  • An access key to one of the victim's AWS accounts was obtained through a weakness in an internet-facing application; the application and the weakness were not named
  • Four access keys belonging to four separate AWS accounts were used from the same source address and client within one second
  • Secrets were collected from AWS Secrets Manager, Systems Manager Parameter Store, plaintext objects in S3, ECS and EC2 environment variables and GitHub and Bitbucket runner environments
  • The actor created IAM users and access keys, opened reverse shells on EC2 instances and ECS containers, and backdoored deployment files, templates and container images
  • Several hundred unique SQL queries ran across dozens of RDS databases for user and transaction data; the first-hand write-up did not confirm that data left the environment
  • Broad compromise came about 72 hours after initial access; the actor then denied access to S3 buckets, set ECS services to zero capacity, purged SQS queues and blocked network access as extortion leverage

Sources

Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.

  1. 1.
    Inside an AI-Assisted Cloud Attack: Familiar Techniques at Unfamiliar SpeedSygnia · Jun 29, 2026 · ResearchEntry vector, key reuse across accounts, secret sources, persistence, data access, impact actions, AI-assistance assessment, duration
  2. 2.
    Threat Actors Uses Agentic AI to Rapidly Compromise Cloud TargetInfosecurity Magazine · Jul 8, 2026 · NewsPress report of the Sygnia findings, with comments from Sygnia's incident-response lead

The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.

More cloud attack breakdowns

All breakdowns