2026-11-17 — Amazon-observed ARTEX AI gateway credential-theft operations

Threat Overview

Amazon Threat Intelligence observed ARTEX agent prompts and operating-memory records reaching its MadPot decoy LLM endpoints from October 3 to November 5. The records described industrial-scale theft of API keys from exposed AI gateways and app builders, validation of cloud credentials, and reuse of stolen model keys to fund further attacks. Targets described included LiteLLM, Dify, Langflow, RAGFlow, Open WebUI, One API and related services. Amazon directly observed the agent-to-decoy requests, but explicitly could not verify every claimed compromise in the agent records. This hunt focuses on actionable credential and cloud-abuse behavior, rather than attributing all suspicious AI API usage to ARTEX.

References

AWS Security Blog, 13 Nov 2026: https://aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents/

Google Threat Intelligence Group, 16 Nov 2026: https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold

Impacted Systems

Self-managed or publicly exposed AI model gateways and application builders: LiteLLM, One API/New API, Dify, Langflow, RAGFlow, FastGPT, Open WebUI and related products; cloud IAM identities and model-service credentials in AWS/Azure. This is a multi-product attack campaign, not a single vulnerability with a common affected version. AWS describes exploit paths including uninitialized installs, default keys, exposed management endpoints, SSRF and previously disclosed product-specific CVEs; versions and patches must be determined per deployed product. Internet-reachable administrative endpoints and exposed credentials increase risk.

Why this matters

A compromised AI gateway can disclose multiple upstream provider tokens and cloud credentials, enabling model abuse, costs and privilege expansion across tenant resources.

Exploitation Status

Confirmed agent requests to Amazon decoys; reported campaign history in prompts is attacker-authored and not independently verified end to end. Google independently documented wider adversarial agentic tooling on 16 November, not necessarily the same operation.

What this hunt looks for

Layered, telemetry-dependent searches for exploit or credential-abuse behavior, followed by source-IP, network, process, identity and persistence pivots. These queries have not been executed in a customer workspace.

Required logs

AWSCloudTrail, AuditLogs, AADServicePrincipalSignInLogs, DeviceProcessEvents, DeviceNetworkEvents. Only hunts whose tables are ingested into Microsoft Sentinel are applicable.

A01 — AWS identity reconnaissance by access key

AWSCloudTrail
| where TimeGenerated >= ago(14d)
| where EventName in~ ("GetCallerIdentity","ListUsers","ListRoles","ListAccessKeys","ListAttachedRolePolicies","ListPolicies")
| summarize Calls=count(), Operations=dcount(EventName), Names=make_set(EventName,20), IPs=make_set(SourceIpAddress,10) by UserIdentityAccessKeyId
| where Operations >= 4
| order by Calls desc

A02 — AWS secrets access bursts

AWSCloudTrail
| where TimeGenerated >= ago(14d)
| where EventName in~ ("GetSecretValue","GetParameter","GetParameters","GetParametersByPath","ListSecrets")
| summarize Calls=count(), Names=make_set(EventName,10), IPs=dcount(SourceIpAddress) by UserIdentityAccessKeyId, bin(TimeGenerated,1h)
| where Calls >= 10
| order by Calls desc

A03 — AWS IAM access key creation after reconnaissance

AWSCloudTrail
| where TimeGenerated >= ago(14d)
| where EventName in~ ("CreateAccessKey","CreateUser","AttachUserPolicy","PutUserPolicy","AssumeRole")
| project TimeGenerated, EventName, UserIdentityArn, UserIdentityAccessKeyId, SourceIpAddress, AWSRegion, ErrorCode
| order by TimeGenerated desc

A04 — AWS Bedrock model usage by unusual identities

AWSCloudTrail
| where TimeGenerated >= ago(14d)
| where EventName has_any ("InvokeModel","Converse","CreateModelInvocationJob")
| summarize Calls=count(), Regions=make_set(AWSRegion,10), SourceIPs=make_set(SourceIpAddress,10) by UserIdentityArn, EventName
| order by Calls desc

A05 — Entra service-principal credential changes

AuditLogs
| where TimeGenerated >= ago(14d)
| where OperationName has_any ("Add service principal credentials","Update application","Add application password","Add service principal")
| project TimeGenerated, OperationName, InitiatedBy, TargetResources, Result
| order by TimeGenerated desc

A06 — Service-principal sign-ins from many IPs

AADServicePrincipalSignInLogs
| where TimeGenerated >= ago(14d)
| summarize SignIns=count(), IPs=dcount(IPAddress), IPList=make_set(IPAddress,20) by ServicePrincipalId, AppId
| where IPs >= 3
| order by IPs desc

A07 — AI gateway server shells and secret-file reads

DeviceProcessEvents
| where TimeGenerated >= ago(14d)
| where FileName in~ ("sh","bash","python","python3","curl","wget","cat")
| where ProcessCommandLine has_any (".env","/proc/self/environ","169.254.169.254","/admin","secrets")
| project TimeGenerated, DeviceName, FileName, ProcessCommandLine, InitiatingProcessFileName, AccountName

A08 — AI application process reaching unusual remote endpoints

DeviceNetworkEvents
| where TimeGenerated >= ago(14d)
| where InitiatingProcessCommandLine has_any ("litellm","langflow","ragflow","dify","open-webui")
| where isnotempty(RemoteIP)
| summarize Connections=count(), Devices=dcount(DeviceId), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated) by InitiatingProcessFileName, RemoteIP, RemoteUrl
| order by Connections desc

Detection Notes

AWSCloudTrail ingestion is necessary for AWS hunts, and only data events enabled in CloudTrail will show relevant data access. Azure/Entra hunts require SigninLogs or AuditLogs. Endpoint queries need Defender for Endpoint DeviceProcessEvents/DeviceNetworkEvents. Legitimate AI workloads can create noisy model API activity; correlate with principal, IP and baseline.

Leave a Reply