Request demo
Back to Blog

6 min read · November 19, 2026

SILENT CLAIMANT: Tracing a Threat Actor Suspected of Using ARTEX to Exfiltrate Data in AWS

How Exaforce linked phishing, scanning and AWS intrusions at unrelated organizations to one actor using stolen workload tokens, long-lived access keys and machine-speed discovery.

A masked figure on an ID badge in front of a cloud, with a key

Introduction

Between late August and October 2026, Exaforce observed phishing, scanning, and intrusions from rented servers on AS197769, directed at several unrelated organizations we monitor. We attribute this activity to a single threat actor that Exaforce tracks as SILENT CLAIMANT (EXA-2026-001). Exaforce worked with contacts at Amazon Web Services to block the subnet associated with this actor, blocking its impact on all AWS infrastructure.

SILENT CLAIMANT gains initial access through internet-facing applications, replaying breached credentials and abusing account-recovery flows on public web portals. In one environment, an unauthenticated endpoint on a third-party application the organization ran itself exposed that application's long-lived IAM access key and the token that gives its Amazon Elastic Kubernetes Service (EKS) pod an AWS identity. In all instances we observed, the primary goal was to exfiltrate data.

These techniques rely on valid credentials and on application features behaving as designed, so each event looks routine in isolation. The activity stood out when we compared identities with their own history across environments.

Given the volume and velocity at which this actor operated, we assess that SILENT CLAIMANT operates an agentic offensive framework such as ARTEX, which chains large language models to automate reconnaissance and exploitation. Tempo and session-naming artifacts support that judgment.

Background: Tracing the Activity to One Network

We linked this activity by correlating activity across multiple customer tenants. A phishing relay, a scanner, and a burst of sign-in attempts in separate environments each looked like routine internet noise, and all came from addresses in 102.220.160.0/22. Searching the rest of the block surfaced activity going back to late August.

Following the block's ownership produces a chain that crosses two regional registries.

text
Attacker-controlled host in 102.220.160.0/22, geolocated to a Frankfurt datacenter
  -> AS197769 (VPSDEDICATED-AS), created in RIPE in May 2026 for a US company
     registered there that February
    -> Sponsoring local internet registry: a German hosting provider present at
       internet exchanges in Amsterdam and Frankfurt
      -> RIPE NCC, the registry for the autonomous system
         Address block: registered with AFRINIC to an internet service provider in Nigeria

AFRINIC space announced through a RIPE autonomous system is a cross-registry mismatch, and that system was under four months old when this activity began. The combination explains why the addresses carried no reputation, and public feeds including AbuseIPDB and VirusTotal held no entries for the ones we checked. We have no indication that the hosting provider or the sponsoring registry is involved in the activity, and a registration record identifies a registrant rather than an operator.

Tactics and Techniques

Initial Access Through Public Web Portals

Where a portal lets anyone register, SILENT CLAIMANT creates its own account and leaves it dormant, likely to give later activity a plausible history.

In one case, the actor leveraged a flaw in recovery code generation: codes were not bound to the requesting account and did not expire. This allowed the attacker to submit values against accounts it does not own. SILENT CLAIMANT used that weakness to take over existing accounts.

Access Through a Third-Party Application

In one environment, scanning found an unauthenticated endpoint on a third-party application the organization ran in its own AWS account. Reaching that endpoint gave access to the application's own environment. That included the IAM roles for service accounts (IRSA) token of its pod, and a long-lived IAM access key in the application's configuration, where such keys are commonly kept.

Stolen Service Account Tokens

Under IRSA, the AWS SDK in a pod exchanges the pod's token for role credentials in an AssumeRoleWithWebIdentity request, which itself needs no AWS credentials. The token is valid for 24 hours by default, and each exchange returns fresh credentials. A copied token therefore keeps producing sessions until it expires, whether or not the pod still exists.

SILENT CLAIMANT sends stolen tokens to AWS Security Token Service (STS) from its own servers, in some cases with Python's urllib library rather than an AWS SDK. In CloudTrail they look like ordinary web identity federation from an unexpected address. The AssumeRoleWithWebIdentity event identifies the service account whose token was used. The role session names, which the caller chooses, pair a one-letter prefix with a three-digit counter and sometimes a task suffix.

EKS Pod Identity narrows this exposure, because its token exchange also requires the node's credentials, provided pods cannot reach the instance metadata service.

Long-Lived Access Keys

SILENT CLAIMANT also converts long-lived IAM user access keys into temporary session credentials with GetSessionToken. Those credentials keep most of the user's permissions for up to 36 hours, even if the key is deactivated. They also carry a new access key ID, so a watchlist built on the original key misses follow-on calls.

Discovery and Machine-Speed Operation

With working credentials, SILENT CLAIMANT checks its identity and permissions, lists S3 buckets, queries AWS Glue, and probes dozens of services, including EC2, ECR, EKS, IAM, RDS, and Secrets Manager, along with cross-account roles and CloudTrail settings. Some calls carry a user agent styled as internal tooling.

Sign-in activity and API discovery arrive in bursts measured in minutes, the source address often stays identical across phases that are hours apart, and the sequence moves from one technique to the next and on into data discovery without pausing for a human decision. The pattern reads as a long-running process on a single rented host rather than an operator at a keyboard.

Possible Use of an ARTEX-like Framework

Shared infrastructure and overlapping timing support our assessment that one threat actor conducted this activity.

Session names that pair an incremental counter with a probe-style suffix, the burst tempo, and one source address persisting across phases resemble artifacts that Google Threat Intelligence Group has published for ARTEX. Several unrelated clusters use that tooling, so the overlap does not associate this activity with any of them.

We have not linked SILENT CLAIMANT to a MITRE ATT&CK group ID or to another vendor's named actor, and organizations outside our visibility may also have seen the same infrastructure.

Where Static Detections Fall Short

Detections that rely on fixed indicators had little to match. The addresses came from a network only months old, IP reputation data from reputable intelligence sources (AbuseIPDB, VirusTotal) was clean, and rotation kept per-address counts low.

A recovered account signs in cleanly and an unauthenticated endpoint returns data by design, so the entry itself produced no failures to count. Data pipelines list buckets every day, so alerts on bucket listing alone are mostly noise. A per-identity threshold on denied calls would catch a discovery burst once it is well under way.

Anomaly-based detections that compare each identity with its own history are likely the best leading indicator of such an actor. A workload role federating from outside its cluster network stands out against that history, as does an access key appearing on an unfamiliar network or a service identity calling services it has never used. That comparison is what surfaced this activity in Exaforce.

Indicators of Compromise

To help the wider community hunt for this activity, we are sharing the indicators below. Rented addresses change hands quickly, so the network indicators will age faster than the behavioral ones. We last observed activity from the block in October 2026, and we will update this post if SILENT CLAIMANT moves to new infrastructure.

Network Indicators

IndicatorTypeDescriptionFirst seen
102.220.160.0/22CIDRBlock used throughout this activity, announced by AS1977692026-08-27
102.220.160.24IPv4Resolution target for a domain of unknown purpose2026-09-25
102.220.160.55IPv4SMTP phishing relay2026-08-27
102.220.160.201IPv4Scanning2026-09-02
102.220.161.64IPv4Scanning2026-10-02
102.220.161.87IPv4Scanning, SMTP relay2026-09-08
102.220.161.102IPv4Scanning2026-09-29
102.220.161.139IPv4Scanning2026-09-26
102.220.163.65IPv4SMTP relay2026-09-29
102.220.163.155IPv4Scanning2026-10-06

Table 1: Network indicators

AS197769 also announces 31.57.37.0/24, 31.57.184.0/24, 31.57.216.0/24, 130.12.181.0/24, and 130.12.182.0/24. We did not observe SILENT CLAIMANT using them, but one address in 31.57.184.0/24 has more than 3,000 public reports of SSH brute forcing. Most organizations can block the entire autonomous system from their sign-in pages and cloud APIs.

Hunting Patterns

These patterns need CloudTrail management read events, because AssumeRoleWithWebIdentity and discovery calls are recorded as read-only. CloudTrail Event history keeps 90 days of management events in each Region, whatever your trail settings. Correlation windows should span several hours.

Log sourceHunt for
CloudTrail, STSAssumeRoleWithWebIdentity where userIdentity.identityProvider contains oidc.eks. and sourceIPAddress sits outside that cluster's egress ranges and VPC CIDRs. Calls through an STS VPC endpoint log private addresses. userIdentity.userName names the service account, and responseElements.credentials.accessKeyId links the session's later calls
CloudTrail, STSGetSessionToken called by an IAM user that does not normally call it
CloudTrail, STSA role session name, the last segment of an assumed-role ARN, matching ^i\d{3}[a-z0-9-]*$, or any naming pattern that role has not used before. boto3 defaults to botocore-session-<epoch>, and other SDKs differ
CloudTrail, any serviceA user agent imitating your internal tooling, from an identity or network that tooling never uses
CloudTrail, any serviceA user agent matching no AWS SDK, CLI, or console format, excluding AWS service principals and AWS Internal
CloudTrail, any serviceAn identity's first cross-account AssumeRole, or its first CloudTrail read (DescribeTrails, GetEventSelectors, LookupEvents)
Kubernetes auditserviceaccounts/token created for an IRSA service account by a caller other than system:node:*, or pods/exec into a pod that runs as one
Application logs with S3 data eventsAn application reporting a failed retrieval where S3 data events record a read of the same object, matched on object key or request ID. Timeouts and retries also produce such pairs
NetworkA third-party application's egress to addresses it has never used
NetworkAny successful sign-in or API call from an AS197769 prefix
NetworkOutbound TCP or UDP 21116 from your hosts to any AS197769 prefix, which is self-hosted RustDesk relay traffic

Table 2: Hunting patterns

User Agents

SubstringContext
X11; Linux x86_64 with Chrome/128.0.0.0Credential stuffing. Match on substrings, since our records may abbreviate the full string
Boto3/1.43.120 with 7.2.11-200.fc44.x86_64 and Python 3.13.12AWS SDK discovery from a Fedora Linux host
aws-cli/2.15.0AWS CLI discovery
Python-urllib/3.13AWS API calls, including STS
curl/8.19.0, python-requests/2.32.5Sign-in and AWS requests; too common to alert on alone

Table 3: User agents

MITRE ATT&CK Mapping

TacticTechnique
Resource DevelopmentT1583.003 Acquire Infrastructure: Virtual Private Server
ReconnaissanceT1595 Active Scanning
Initial AccessT1190 Exploit Public-Facing Application
Initial AccessT1566 Phishing
Credential AccessT1110.004 Brute Force: Credential Stuffing
Credential AccessT1528 Steal Application Access Token
Lateral MovementT1550.001 Use Alternate Authentication Material: Application Access Token
Initial Access, PersistenceT1078.004 Valid Accounts: Cloud Accounts
DiscoveryT1526 Cloud Service Discovery
DiscoveryT1580 Cloud Infrastructure Discovery
DiscoveryT1087.004 Account Discovery: Cloud Account
DiscoveryT1069.003 Permission Groups Discovery: Cloud Groups
DiscoveryT1613 Container and Resource Discovery
DiscoveryT1654 Log Enumeration
StealthT1036 Masquerading

Table 4: MITRE ATT&CK mapping. Tactic names follow the current matrix, which renamed TA0005 from Defense Evasion to Stealth.