AI-Driven Identity and Cloud Exploitation
An adversary is using an AI-orchestration framework to conduct high-velocity credential spraying against Entra ID and subsequently manipulate application credentials or directory roles for persistence at machine speed.
Based on research by Cisco Talos 2026-11-14 9 steps · 3 queries T1078.004 T1098 T1110.003 T1550.001
Brief
Why Now
Recent research from Cisco Talos in their report ARTEX and beyond: AI-enabled penetration testing platforms details the rise of frameworks that orchestrate attacks at machine speed. These tools allow even low-skill actors to execute complex identity exploitation chains. They transition from initial credential spraying to cloud persistence in minutes, often outpacing standard detection logic that looks for isolated events. We designed this hunt to find the automated fingerprints of these AI-driven campaigns within Entra ID environments.
How the Hunt Flows
The first phase focuses on the hb_auth_signin surface. A query identifies source IPs that target a high volume of unique accounts with authentication failures in under one hour. We filter for automated patterns by examining the ratio of accounts targeted to the time window. This identifies the infrastructure used for the initial spray, whether it originates from known data centers or suspicious geographic locations.
The hunt then fans out to inspect two secondary surfaces for signs of follow-on persistence. On the hb_cloud_api_activity surface, we look for sensitive Microsoft Graph operations such as adding application passwords or updating credentials. Simultaneously, the hunt checks the hb_entity_management surface for rare or unexpected directory role assignments, focusing on high-privilege roles like Global Administrator or Application Administrator.
A correlation agent then bridges these findings. It evaluates the temporal proximity between the spraying IPs and the administrative changes. The analyst or agent determines if a targeted account was used to perform an administrative API call shortly after the spray attempt. If a link exists, the hunt provides automated actions to revoke sessions and rotate credentials.
Blind Spots
Log latency is the primary risk for this hunt. Microsoft 365 Unified Audit Logs can have significant ingestion delays. An AI-driven framework may complete its entire objective before the logs reach the telemetry surface, making this a reactive rather than proactive control. Additionally, if an attacker uses a highly distributed residential proxy network, the per-IP threshold used in the first query might fail to trigger. In such cases, the hunt relies more heavily on the persistence-based queries rather than the initial spray leads.
In this series
Steps
-
High-velocity credential spraying IPs
Query · detection candidateIdentify source IPs that target a large number of unique accounts with authentication failures in under one hour, suggesting the use of an automated framework like ARTEX.
reads hb_auth_signinsqlSELECT src_endpoint_ip, user_agent, src_location_country, COUNT(DISTINCT actor_user_name) AS unique_accounts_targeted, COUNT(*) AS failure_count, MIN(time) AS first_failure, MAX(time) AS last_failure, (julianday(MAX(time)) - julianday(MIN(time))) * 24 AS duration_hours FROM hb_auth_signin WHERE status_id = 2 AND provider IN ('azure_ad', 'm365') AND time >= datetime('now', '-{{lookback_days}} days') GROUP BY src_endpoint_ip, user_agent, src_location_country HAVING unique_accounts_targeted >= {{spray_threshold}} AND duration_hours < 1.0 ORDER BY unique_accounts_targeted DESCWhat a hit looks like. Source IPs targeting multiple accounts in a burst under one hour. The user agent and country help distinguish automated infrastructure from legitimate corporate clusters.
-
Application credential modifications
Query · enrichmentFind instances where an actor added application credentials to establish persistence after a successful spray.
reads hb_cloud_api_activitysqlSELECT time, actor_user_name, api_operation, resource_name, src_endpoint_ip, status FROM hb_cloud_api_activity WHERE provider = 'azure_ad' AND instr(',' || '{{sensitive_api_ops}}' || ',', ',' || api_operation || ',') > 0 AND time >= datetime('now', '-{{lookback_days}} days')What a hit looks like. Rows indicate where an actor added application credentials such as passwords or keys. Correlating these actors with the spraying IPs provides high-fidelity evidence.
-
Privileged role assignments
Query · baselineIdentify accounts granted high-privilege directory roles, tracking both standard and non-standard assignment operations.
reads hb_entity_managementsqlSELECT entity_name AS role_name, actor_user_name, operation_type, time, device_hostname FROM hb_entity_management WHERE provider = 'azure_ad' AND activity_id IN (1, 3, 99) AND instr(',' || '{{privileged_roles}}' || ',', ',' || entity_name || ',') > 0 AND time >= datetime('now', '-{{lookback_days}} days')What a hit looks like. Unexpected grants of Global Admin or similar roles. AI frameworks typically move to this stage rapidly after breaching an account.
-
Triage AI-driven identity threats
Agent triageAnalyze the correlation between high-speed spraying IPs and subsequent privileged operations to confirm a machine-speed intrusion.
-
Route on verdict
DecisionImmediately route to containment if an automated identity attack is confirmed.
-
Revoke sessions and credentials
Response actionImmediately contain the breach by revoking tokens and rotating credentials for involved accounts.
-
Incident follow-up and investigation
Analyst taskConduct a full investigation into the scope of the intrusion and verify remediation.
-
Hunt closure and reporting
Analyst taskRecord findings and update defensive controls.
Coverage
Scenario coverage
| Stage | Covered | How, or why not |
|---|---|---|
| High-Velocity Credential Spraying T1110.003 |
Yes | lead-spraying-ips |
| Cloud Privilege Manipulation T1098 |
Yes | app-credential-changes, role-assignment-changes |
| Automated Vulnerability Probing T1190 |
Out of scope | Belongs to another part of the 'ARTEX and beyond: AI-enabled penetration testing platforms' series. |
| Orchestrated C2 and Tunneling T1572 |
Out of scope | Belongs to another part of the 'ARTEX and beyond: AI-enabled penetration testing platforms' series. |
| Internal Discovery Automation T1046 |
Out of scope | Belongs to another part of the 'ARTEX and beyond: AI-enabled penetration testing platforms' series. |
| Automated Credential Harvesting T1003.001 · T1555 |
Out of scope | Belongs to another part of the 'ARTEX and beyond: AI-enabled penetration testing platforms' series. |
Blind spots
- Needs low-latency telemetry for M365 Unified Audit Log. A machine-speed framework may complete its objective before the logs are available for the hunt, leading to a delayed response. It would answer whether the spraying happened in the last 60 minutes. Remediation: Enable high-priority log streaming for Entra ID audit events.
- Needs hb_auth_signin with is_proxy data. If the attacker uses residential proxies as reported for Hephaestus, a single IP threshold might miss a distributed spray. It would answer which IPs are residential proxies used to bypass per-IP thresholds. Remediation: Review accounts with failures from multiple 'New Geo-Location' behaviors in Okta/Entra data.
Parameters & data
Parameters
| Parameter | Type | Default | What it is |
|---|---|---|---|
lookback_days | number | 14 | Days of history to examine for identity activity. |
privileged_roles | list[string] | Global Administrator, Application Administrator, Cloud Application Administrator, Privileged Role Administrator | High-privilege Entra ID roles to monitor for unauthorized assignments. |
sensitive_api_ops | list[string] | Add application password, Add key credential, Update application, Add service principal | API operations indicating application credential manipulation in Entra ID. |
spray_threshold | number | 10 | Minimum number of unique accounts targeted by a single IP to qualify as a spray lead. |
Telemetry
| Source | Category | Telemetry |
|---|---|---|
| Endpoint telemetry (hb_ surfaces) | endpoint | endpoint |
| Identity / sign-in telemetry | identity | identity |
Source
---
analysis: A standard rule detects individual password sprays; this hunt correlates
the spray with Graph API manipulation and rare role assignments across three surfaces,
providing the context to identify an orchestrated AI intrusion.
blind_spots:
- id: m365-audit-latency
owner: Cloud Infrastructure Team
question: whether the spraying happened in the last 60 minutes
remediation: Enable high-priority log streaming for Entra ID audit events.
requires: low-latency telemetry for M365 Unified Audit Log
risk: A machine-speed framework may complete its objective before the logs are available
for the hunt, leading to a delayed response.
stage: high-velocity-credential-spraying
- id: residential-proxy-evasion
owner: Security Operations
question: which IPs are residential proxies used to bypass per-IP thresholds
remediation: Review accounts with failures from multiple 'New Geo-Location' behaviors
in Okta/Entra data.
requires: hb_auth_signin with is_proxy data
risk: If the attacker uses residential proxies as reported for Hephaestus, a single
IP threshold might miss a distributed spray.
stage: high-velocity-credential-spraying
coverage:
- stage: high-velocity-credential-spraying
status: covered
steps:
- lead-spraying-ips
- stage: cloud-privilege-manipulation
status: covered
steps:
- app-credential-changes
- role-assignment-changes
- reason: 'Belongs to another part of the ''ARTEX and beyond: AI-enabled penetration
testing platforms'' series.'
stage: automated-vulnerability-probing
status: out_of_scope
- reason: 'Belongs to another part of the ''ARTEX and beyond: AI-enabled penetration
testing platforms'' series.'
stage: orchestrated-c2-and-tunneling
status: out_of_scope
- reason: 'Belongs to another part of the ''ARTEX and beyond: AI-enabled penetration
testing platforms'' series.'
stage: internal-discovery-automation
status: out_of_scope
- reason: 'Belongs to another part of the ''ARTEX and beyond: AI-enabled penetration
testing platforms'' series.'
stage: automated-credential-harvesting
status: out_of_scope
guardrails:
claims: no_unsupported
evidence: citation_required
missing_data: not_benign
telemetry: untrusted
hunt:
applicability: campaign-specific
handoff: promote-to-detection
justification: AI-enabled frameworks allow attackers to scale identity compromises
across thousands of tenants at machine speeds; a negative result over the estate
confirms that automated persistence tradecraft is not active.
methodology: model-assisted
trigger: intel-report
hypothesis: An adversary is using an AI-orchestration framework to conduct high-velocity
credential spraying against Entra ID and subsequently manipulate application credentials
or directory roles for persistence at machine speed.
labels:
- hunt
- attack.t1110.003
- attack.t1098
- attack.t1550.001
- attack.t1078.004
- command and control
- credential access
- discovery
- initial access
- persistence
name: AI-Driven Identity and Cloud Exploitation
parameters:
lookback_days:
default: '14'
description: Days of history to examine for identity activity.
from:
kind: manual
observed: '2026-11-13'
ref: standard-retention
type: number
privileged_roles:
default:
- Global Administrator
- Application Administrator
- Cloud Application Administrator
- Privileged Role Administrator
description: High-privilege Entra ID roles to monitor for unauthorized assignments.
from:
kind: article
observed: '2026-11-13'
ref: talos-artex-report
type: list[string]
sensitive_api_ops:
default:
- Add application password
- Add key credential
- Update application
- Add service principal
description: API operations indicating application credential manipulation in
Entra ID.
from:
kind: article
observed: '2026-11-13'
ref: talos-artex-report
type: list[string]
spray_threshold:
default: '10'
description: Minimum number of unique accounts targeted by a single IP to qualify
as a spray lead.
from:
kind: article
observed: '2026-11-13'
ref: talos-artex-report
type: number
provenance:
authors:
- name: Huntbase hunt generation
org: huntbase.io
generated:
by: huntbase-hunt-generation
from: https://blog.talosintelligence.com/artex-and-beyond/
gates:
- dry-run
- lint
model: hb_google/gemini-3-flash-preview
rationale: Focus on Entra ID (Azure AD) and Microsoft 365 tenants. High-velocity spraying
is best identified on the sign-in surface, while persistence moves appear in cloud
API and entity management logs.
references:
- name: 'ARTEX and beyond: AI-enabled penetration testing platforms'
url: https://blog.talosintelligence.com/artex-and-beyond/
related:
- hunt: automated-vulnerability-probing
reason: This hunt focuses on identity and cloud persistence; external vulnerability
exploitation requires hb_http_activity and is handled separately.
relation: out-of-scope-alternative
scenario:
stages:
- name: Automated Vulnerability Probing
observables:
- Acunetix web scanner
- Nessus vulnerability scanner
- Interactsh out-of-band detection
- Inbound scanning traffic from ARTEX-associated C2 IPs
slug: automated-vulnerability-probing
tactic: initial-access
techniques:
- T1190
- name: High-Velocity Credential Spraying
observables:
- Accelerated sign-in attempts against downstream Microsoft 365 tenants
- Compromised ConnectWise application credentials used for multi-tenant access
slug: high-velocity-credential-spraying
tactic: credential-access
techniques:
- T1110.003
- name: Orchestrated C2 and Tunneling
observables:
- VShell C2 framework
- AdaptixC2 framework
- Chisel TCP/UDP tunneling
- Net Proxy Server (NPS) tunneling
- multios.trojan.adaptixc2 signature
slug: orchestrated-c2-and-tunneling
tactic: command-and-control
techniques:
- T1572
- name: Internal Discovery Automation
observables:
- fscan internal network scanner
- multios.tool.fscan signature
slug: internal-discovery-automation
tactic: discovery
techniques:
- T1046
- name: Automated Credential Harvesting
observables:
- Mimikatz execution
- LSASS memory access
- Credential leak management services integration
slug: automated-credential-harvesting
tactic: credential-access
techniques:
- T1003.001
- T1555
- name: Cloud Privilege Manipulation
observables:
- Microsoft Graph API activity
- Entra ID application credential changes
- Role assignment and privilege modifications
slug: cloud-privilege-manipulation
tactic: persistence
techniques:
- T1098
summary: Threat actors are leveraging AI-driven penetration testing frameworks like
ARTEX and PentAGI to automate reconnaissance, credential spraying, and post-exploitation
at machine speed. These frameworks orchestrate a variety of offensive tools including
C2 frameworks (VShell, AdaptixC2), tunneling utilities (Chisel, NPS), and internal
scanners (fscan) to rapidly compromise targets across both endpoint and cloud
environments.
series:
index: 1
slug: artex-and-beyond-ai-enabled-penetration-testing-platforms
title: 'ARTEX and beyond: AI-enabled penetration testing platforms'
total: 2
severity: high
targets:
analyst:
name: Tier-2 analyst
role: analyst
endpoint:
category: endpoint
name: Endpoint telemetry (hb_ surfaces)
telemetry:
- endpoint
hunter:
agent: true
name: Hunt agent
identity:
category: identity
name: Identity / sign-in telemetry
telemetry:
- identity
tlp: clear
type: investigation
---
# AI-Driven Identity and Cloud Exploitation
This hunt targets the high-speed identity exploitation patterns observed in ARTEX campaigns, specifically the automation of credential spraying and follow-on cloud control-plane manipulation. It begins by identifying source IPs exhibiting automated sign-in failure patterns across many accounts in a short window. The hunt then fans out to search for correlated evidence of persistence, such as an actor adding application passwords via Microsoft Graph or unauthorized directory role assignments. An agent evaluates the temporal and actor-based links between the spraying activity and the administrative changes to confirm an AI-driven intrusion.
## lead-spraying-ips
<!-- High-velocity credential spraying IPs -->
Identify source IPs that target a large number of unique accounts with authentication failures in under one hour, suggesting the use of an automated framework like ARTEX.
```sqlite target=identity role=detection-candidate params=(lookback_days=lookback_days, spray_threshold=spray_threshold)
~~~yaml
expected: Source IPs targeting multiple accounts in a burst under one hour. The user
agent and country help distinguish automated infrastructure from legitimate corporate
clusters.
reads:
- actor_user_name
- provider
- src_endpoint_ip
- src_location_country
- status_id
- time
- user_agent
silence: not_evidence_of_absence
source: hb_auth_signin
verified: dry-run
verified_at: '2026-11-14'
~~~
SELECT src_endpoint_ip, user_agent, src_location_country, COUNT(DISTINCT actor_user_name) AS unique_accounts_targeted, COUNT(*) AS failure_count, MIN(time) AS first_failure, MAX(time) AS last_failure, (julianday(MAX(time)) - julianday(MIN(time))) * 24 AS duration_hours FROM hb_auth_signin WHERE status_id = 2 AND provider IN ('azure_ad', 'm365') AND time >= datetime('now', '-{{lookback_days}} days') GROUP BY src_endpoint_ip, user_agent, src_location_country HAVING unique_accounts_targeted >= {{spray_threshold}} AND duration_hours < 1.0 ORDER BY unique_accounts_targeted DESC
```
## investigate-fan-out
<!-- Fan out for cloud persistence -->
parallel:
- → app-credential-changes
- → role-assignment-changes
join: → agent-correlation
## app-credential-changes
<!-- Application credential modifications -->
Find instances where an actor added application credentials to establish persistence after a successful spray.
```sqlite target=endpoint role=enrichment params=(lookback_days=lookback_days, sensitive_api_ops=sensitive_api_ops)
~~~yaml
expected: Rows indicate where an actor added application credentials such as passwords
or keys. Correlating these actors with the spraying IPs provides high-fidelity evidence.
reads:
- actor_user_name
- api_operation
- provider
- resource_name
- src_endpoint_ip
- status
- time
silence: not_evidence_of_absence
source: hb_cloud_api_activity
verified: dry-run
verified_at: '2026-11-14'
~~~
SELECT time, actor_user_name, api_operation, resource_name, src_endpoint_ip, status FROM hb_cloud_api_activity WHERE provider = 'azure_ad' AND instr(',' || '{{sensitive_api_ops}}' || ',', ',' || api_operation || ',') > 0 AND time >= datetime('now', '-{{lookback_days}} days')
```
## role-assignment-changes
<!-- Privileged role assignments -->
Identify accounts granted high-privilege directory roles, tracking both standard and non-standard assignment operations.
```sqlite target=endpoint role=baseline params=(lookback_days=lookback_days, privileged_roles=privileged_roles)
~~~yaml
baseline:
compare: new_this_window
window: '{{lookback_days}}d'
expected: Unexpected grants of Global Admin or similar roles. AI frameworks typically
move to this stage rapidly after breaching an account.
prevalence:
by: actor_user_name
key:
- entity_name
rare_below: 3
reads:
- activity_id
- actor_user_name
- device_hostname
- entity_name
- operation_type
- provider
- time
silence: not_evidence_of_absence
source: hb_entity_management
verified: dry-run
verified_at: '2026-11-14'
~~~
SELECT entity_name AS role_name, actor_user_name, operation_type, time, device_hostname FROM hb_entity_management WHERE provider = 'azure_ad' AND activity_id IN (1, 3, 99) AND instr(',' || '{{privileged_roles}}' || ',', ',' || entity_name || ',') > 0 AND time >= datetime('now', '-{{lookback_days}} days')
```
## agent-correlation
<!-- Triage AI-driven identity threats -->
```agent target=hunter
cite: required
context:
- lead-spraying-ips
- app-credential-changes
- role-assignment-changes
max_iterations: 5
objective: Determine if any IPs or accounts from the spraying leads correlate with
the application credential or role modifications. Look for temporal proximity and
whether a targeted account was subsequently used for an administrative API call.
success_criteria: A per-IP and per-account verdict of malicious or suspicious, citing
the relevant rows from all three surfaces.
tools:
- endpoint
- identity
```
## route-on-result
<!-- Route on verdict -->
if~: "the agent verdict is malicious for at least one IP or account, indicating a successful spray and follow-on administrative change" (confidence: high, judge=hunter)
then: → revoke-sessions
indeterminate: → incident-follow-up
unavailable: → incident-follow-up (blind_spot: m365-audit-latency)
else: → hunt-closure
## revoke-sessions
<!-- Revoke sessions and credentials -->
```action target=identity
~~~yaml
approval: required
~~~
Revoke all active sessions and OAuth tokens for compromised accounts. Disable the affected application registrations and force a password reset for any involved users.
```
→ incident-follow-up
## incident-follow-up
<!-- Incident follow-up and investigation -->
```manual target=analyst
Review all Graph API logs for the compromised accounts to ensure no other applications were backdoored. Verify that the revoked credentials and roles have not been re-instated.
```
→ hunt-closure
## hunt-closure
<!-- Hunt closure and reporting -->
```manual target=analyst
Document the identified spraying IPs and persistence patterns. If a detection candidate query was effective, recommend it for promotion to a standing alert with a threshold adjusted for normal tenant noise.
```
→ end
Run it
Take this hunt into your environment.
Open it in Huntbase to run every step against your own connections, with Scout weighing the evidence and your analysts in command. Or take the open hunt.md file anywhere that reads the format.
Machine-drafted by huntbase-hunt-generation using hb_google/gemini-3-flash-preview, gated by dry-run, lint, then reviewed by a person.