Skip to main content
Back
IntrusionAWS · Azure · Microsoft Entra · SaaS
18 min read

FBI-tracked ARTEX campaigns: AI agents across AWS, Azure and Entra

Victim
Many unnamed organizations across sectors: AWS and Azure customers and Microsoft Entra tenants among them
When
August 2026Disclosed Nov 13, 2026
Entry
Internet-exposed AI gateways and applications (default administrator credentials, unauthenticated configuration endpoints, known CVEs, secrets in client-side files) and cloud keys or application secrets in public code; how the AWS and Azure credentials of the GTIG engagements were obtained is unstated.

ARTEX is an open-source penetration testing framework, written in Go, that turns a goal written in plain language into commands that a team of AI agents plans and runs, with little human input after the goal is set. It has no model of its own: it calls outside large language model providers, and on the instances that attackers ran, low-cost Chinese-origin models such as DeepSeek were the most common. It was released on GitHub in July 2026 for authorized testing. The developer took the repository down after reports of abuse (on October 8 per the FBI; Check Point dates the deletion to September 8), but forks, archived copies and a commercial version keep it in circulation, and the FBI cites over 1,600 consoles exposed on the internet in early November.

The FBI, Amazon, Google Threat Intelligence Group and Mandiant, Microsoft, Check Point and Wiz reported in October and November 2026 that several actors had used ARTEX against many unnamed organizations since August 2026. Its agents searched public code and internet-exposed AI gateways for keys, tested stolen AWS and Azure credentials in seconds, signed in as Entra service principals, read Key Vault secrets, ran commands through Kudu, used Run Command on Azure virtual machines with no result stated, made a vendor app Global Administrator, launched EC2 footholds, staged EBS snapshots and overwrote a public S3 bucket; a password spray, EC2 launches with high-value IAM roles and a cross-tenant read through a vendor application failed. The FBI and Microsoft put the time from initial targeting to privileged access under three hours in multiple instances, while GTIG puts one cluster's move to deep system access at around four hours; exfiltration from the AWS staging is not confirmed, and Amazon could not verify the outcomes in the agents' own records.

This review covers a campaign across many unnamed victims, not one victim's estate. It judges each control as deployed on your own cloud: a row says what would happen if you run that cloud and have that control, and its caveat names the conditions, such as the cloud that hosts the gateway.

What happened

External Parties
ARTEX operators
Model providers
Amazon decoy sensors
Internet
Cloud APIs and Entra IDkeys, tokens, app secrets
AWS Accounts
Azure Subscriptions
+ keys in public code
+ exposed AI gateways
+ upstream provider keys
Public Code and AI Apps
+ Microsoft 365 accounts
+ workload and vendor apps
+ privileged directory roles
Microsoft Entra Tenants
Compromised IAM Users
Vendor Cognito Pools
S3 Landing-Page Bucket
EC2 and EBS Storage
Storage Accounts
App Service and Kudu
Key Vaults
Virtual Machines

Replaying the ARTEX agent campaigns, 2026

  1. 1

    Keys from public code and exposed AI apps

    From August 2026

    Agents search public code for leaked keys and secrets, take upstream provider keys from internet-exposed AI gateways by their own records, and aim forged requests at a cloud metadata service that no source names.

  2. 2

    Testing stolen cloud credentials

    Late August to October 2026

    Agents test compromised AWS IAM user keys with DryRun calls, map Azure keys and tokens in seconds, spray passwords at Microsoft 365 accounts without success, and attempt Amazon Cognito SignUp calls with no stated outcome. In a separate proprietary application flow, registration succeeds but the vendor's tenant check refuses another company's data.

  3. 3

    Identities, secrets and commands in Entra and Azure

    October 2026

    Service principals sign in with stolen application credentials, vendor apps sign in to customer tenants where they were already consented, and the agents read Key Vault secrets, run commands through Kudu and Run Command, add app credentials and make a vendor app Global Administrator.

  4. 4

    Footholds and staging in AWS

    Dates not stated; reported October 30 and November 5, 2026

    Launches with high-value IAM roles fail, but agents set up an EC2 jump host, stage EBS snapshots on keyless instances with exfiltration not confirmed, and overwrite a public S3 bucket with leaked credentials.

  5. 5

    Inference on someone else's bill

    October 3 to November 5, 2026, in Amazon's records

    Agents grade stolen keys by working calls and spendable balance and feed them into the framework's own model gateway; no source names the providers or the spend.

Impact

Across many unnamed victims, ARTEX agents searched public code and exposed AI gateways for keys, aimed forged requests at a cloud metadata service that no source names, tested stolen AWS and Azure credentials, attempted to create accounts through Amazon Cognito SignUp, signed in as Microsoft Entra service principals, targeted Resource Manager operations that return credentials, read Azure Key Vault secrets, ran commands through Kudu, made a vendor app Global Administrator, launched EC2 footholds, staged EBS snapshots and overwrote objects in one public S3 bucket. Microsoft said the actor attempted to validate stolen Azure AI resources, and Amazon assesses that stolen provider keys likely paid for the agents' inference. Password sprays against Microsoft 365 accounts and against Okta and Entra accounts at software vendors, EC2 launches with high-value IAM roles, a security group creation and a cross-tenant read through a vendor application failed. Exfiltration from the AWS staging is not confirmed, and Amazon could not verify the outcomes in the agents' own records.

Outcome

Where cloud-provider controls fit

This page reviews a campaign across many unnamed victims, so it judges each control as deployed on your own cloud: a row says what would happen if you run that cloud and have that control, and its caveat names the conditions. On those terms thirteen rows would have stopped a step outright. A security group on AWS or a network security group on Azure that admits only your own address ranges refuses the agents' requests from the internet to an AI gateway's management interface. IMDSv2 on AWS and the metadata header requirement on Azure refuse a metadata query sent through an XML external entity; the sources do not say what the forgery through URL-fetch features could send. A network perimeter service control policy refuses the DryRun probes of a compromised IAM user from outside your networks, in a member account. Amazon Cognito with self-service sign-up off refuses the public SignUp call. A service control policy that keeps Amazon Bedrock invocations to your AI workloads' networks, and local authentication off on Azure AI accounts, refuse model calls with a stolen key. In Microsoft Entra ID, workload identity Conditional Access refuses the token requests for your own single-tenant apps, which Microsoft places primarily at two addresses, one of them a Tor exit node. A Key Vault firewall, or a resource control policy that keeps AWS Secrets Manager to your networks, refuses a read of a secret with a stolen credential from outside those networks. Access restrictions on the Kudu site of App Service refuse a Kudu request from outside your address ranges. A service control policy that denies ec2:RunInstances and volume creation to all but approved administrators refuses the keyless instances and the new 100 GB volume, if the account is a member account of your organization and the compromised identity is not one of your approved administrators. Two rows stay contained because the sources leave open a fact about the step itself: which credential signed the IAM user probes, for the service control policy that refuses IAM user calls without MFA, and which credential ran the Resource Manager validate sweep, for workload identity Conditional Access. The other cloud rows bound a step. In AWS, VPC Block Public Access, a KMS key policy and security groups with no outbound rule would have bounded the jump host and the snapshot staging. In Azure, storage key access off, a storage account firewall, App Service basic authentication off and an Azure Policy deny on Run Command resources would have bounded what the obtained credentials reached. A vendor's multitenant app in your tenant, and the roles it already holds, stay outside workload identity Conditional Access and the application management policy. Keys in public code, flaws in AI gateway software and keys of model providers outside the clouds stay with their own layers.

Could this have been prevented?

In hindsight, not as a whole. 13 controls would have refused a step outright, and 18 controls would have limited one. 7 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.

13
Prevented
18
Contained
7
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.

External Parties
ARTEX operators
Model providers
Amazon decoy sensors
Internet
Cloud APIs and Entra IDkeys, tokens, app secrets
AWS Accounts
Azure Subscriptions
+ keys in public code
+ exposed AI gateways
+ upstream provider keys
Public Code and AI Apps
+ Microsoft 365 accounts
+ workload and vendor apps
+ privileged directory roles
Microsoft Entra Tenants
Compromised IAM Users
Vendor Cognito Pools
S3 Landing-Page Bucket
EC2 and EBS Storage
Storage Accounts
App Service and Kudu
Key Vaults
Virtual Machines

Replaying the ARTEX agent campaigns, 2026

1 · Keys from public code and exposed AI apps

4 Prevented1 Contained2 Not covered2 on other clouds

2 · Testing stolen cloud credentials

2 Prevented7 Contained1 Not covered

3 · Identities, secrets and commands in Entra and Azure

4 Prevented4 Contained2 Not covered6 on other clouds

4 · Footholds and staging in AWS

1 Prevented6 Contained1 Not covered

5 · Inference on someone else's bill

2 Prevented1 Not covered2 on other clouds

Key facts

  • The FBI said a variety of actors used ARTEX against organizations in many sectors since at least August 2026; the developer made the repository unavailable on October 8, 2026, and forks, archived copies and a commercial version remained.
  • Amazon's decoy sensors, posing as LLM inference APIs, captured ARTEX agents' campaign records from October 3 to November 5, 2026 that covered hundreds of targets; Amazon could not verify the outcomes the agents reported and notified potentially affected AWS customers.
  • GTIG and Mandiant saw agents probe AWS permissions with 27 events in 23 seconds of active burst time in one operation, map Azure credentials with a Resource Manager validation sweep of 11 resource types in 8 seconds, and deploy a 17-snapshot EC2 instance in about 10 seconds; launches with high-value IAM roles seemed to fail on invalid instance profile ARNs.
  • From October 10 to 14, 2026 a Storm-0940 password spray against 327 accounts at eight organizations produced more than 151,000 failed logons, with a peak of 2,900 per minute on October 11; lockouts were nearly 98% of the activity and no authentication succeeded.
  • Microsoft said Storm-0940 attempted sign-ins to more than 2,250 tenants through ConnectWise applications as of October 21, 2026, and the FBI wrote that the compromise enabled access to more than 2,250 downstream tenants; one ConnectWise application self-assigned the Global Administrator role, and a compromised multi-tenant app read secrets at scale across Azure Key Vaults.
  • Check Point counted 1,608 live ARTEX consoles, 2,557 seen in all, as of November 5, 2026; in one campaign leaked S3 credentials let an agent overwrite a public bucket of iOS landing pages, which its owners later restored; the source does not establish whether users received poisoned content. Outside the cloud, lateral movement to domain controllers was blocked, the same crypto campaign moved no funds, and deepfakes against identity verification at two exchanges were rejected in every recorded attempt.

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.
    Threat Overview of ARTEX Autonomous Penetration Testing FrameworkFederal Bureau of Investigation (FLASH-20261113-001) · Nov 13, 2026 · AdvisoryCross-campaign TTP table, cloud permission mapping, cloud persistence, cloud collection, sectors, timing claim, exposure count, mitigations
  2. 2.
    Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agentsAmazon Threat Intelligence · Nov 13, 2026 · ResearchTarget discovery, exposed AI gateway exploitation, provider key theft and validation, cloud expansion, key grading, inference paid by victims, mitigations
  3. 3.
    ARTEX and Friends: Enter Agentic Offensive FrameworksGoogle Threat Intelligence Group and Mandiant · Oct 30, 2026 · ResearchAzure permission mapping, AWS permission testing, Kudu command execution, failed role escalation, AWS persistence, EBS snapshot harvesting preparation, burst timing, defensive guidance
  4. 4.
    AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromiseMicrosoft Threat Intelligence · Oct 21, 2026 · ResearchPassword spray and lockouts, OAuth application credential abuse, credential additions, role self-assignment, Key Vault secret reads, time to privileged access
  5. 5.
    Threat actors use agentic AI tooling to find and abuse exposed secretsMicrosoft Threat Intelligence · Nov 5, 2026 · ResearchGitHub secret searches, service principal sign-ins, Resource Manager enumeration, credential-returning operations, Key Vault exfiltration, Run Command, AI resource validation
  6. 6.
    Hacking on Autopilot: Inside ARTEX Cybercrime CampaignsCheck Point Research · Nov 5, 2026 · ResearchS3 credential leak and bucket overwrite, campaign outcomes, console exposure counts, model costs
  7. 7.
    ARTEX in the Wild: Storm-0940 Expands Targeting of IT and Software VendorsWiz Research · Oct 22, 2026 · ResearchPassword spraying in Okta and Entra, Cognito sign-up attempts, application tenant isolation that held

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