Agentic Identity and Cloud Control Plane Exploitation
An intruder is using an agentic framework to automate cloud API interactions and identity access, characterized by successful sign-ins from unmanaged devices followed by high-tempo API bursts and local CLI tool usage.
Based on research by Elastic Security Labs 2026-11-14 9 steps · 3 queries T1059 T1078 T1090.003 T1505.003
Brief
Why now
Offensive security researchers recently detailed how agentic AI frameworks automate the decision-making process for attackers. In the article "Inside the loop: Five AI attack frameworks that never ask permission" (https://www.elastic.co/security-labs/threat-command/agentic-offensive-frameworks-detection-artex), the authors describe how tools like ARTEX and CyberStrikeAI issue API commands in bursts of 2 to 25 seconds. This speed defeats manual triage. We published this hunt to find the specific footprint of these frameworks in cloud environments.
How the hunt flows
The first phase identifies successful authentications to platforms like AWS, GitHub, or Entra ID that originate from unmanaged or unknown devices. This query scopes the hunt to identities that lack a verified device posture, providing the initial leads for potential session abuse.
The second phase runs two searches in parallel. One query scans cloud API logs for rare, non-browser User-Agents such as Boto3, Python-requests, or the AWS CLI. It filters for agents seen on very few cloud accounts to highlight framework defaults rather than enterprise-wide scripts. Simultaneously, another query checks the hosts identified in the first phase for the execution of cloud management tools or scripting interpreters like Terraform and Python.
Finally, the hunt uses an agent to synthesize the timing of these events. The agent evaluates whether the API bursts and local processes began immediately after the unmanaged sign-in. This temporal correlation helps distinguish a human developer from an automated offensive agentic loop.
What this hunt cannot see
This hunt relies on management events. If an attacker performs high-tempo data exfiltration inside an S3 bucket, it remains invisible unless the environment captures detailed CloudTrail Data Events. Furthermore, if the framework runs inside a short-lived container where no endpoint agent is present, the hunt loses visibility into the local process activity, leaving only the cloud API logs for analysis.
In this series
Steps
-
Cloud sign-ins from unmanaged devices
Query · scopingIdentify successful authentications to critical cloud platforms originating from unmanaged devices or unknown workstations, providing the lead for session abuse.
reads hb_auth_signinsqlSELECT actor_user_name, src_endpoint_ip, src_endpoint_hostname, device_hostname, provider, device_trust_type, time FROM hb_auth_signin WHERE activity_id = 1 AND status_id = 1 AND (device_is_compliant = 'false' OR device_trust_type IS NULL OR device_trust_type = '') AND instr(',' || '{{target_providers}}' || ',', ',' || LOWER(provider) || ',') > 0 AND time >= datetime('now', '-{{lookback_days}} days')What a hit looks like. A list of users and source hosts that signed into cloud services without a managed device context. Silence suggests all sign-ins were from trusted devices.
-
Rare automated User-Agents in cloud API logs
Query · baselineDetect non-browser tools like Boto3 or Python-requests that are rare across the environment, which often facilitate agentic framework interactions.
reads hb_cloud_api_activitysqlSELECT user_agent, actor_user_name, provider, COUNT(DISTINCT cloud_account_uid) AS accounts, COUNT(*) AS calls, MIN(time) AS first_seen FROM hb_cloud_api_activity WHERE (LOWER(user_agent) LIKE '%python%' OR LOWER(user_agent) LIKE '%boto%' OR LOWER(user_agent) LIKE '%aws-cli%' OR LOWER(user_agent) LIKE '%curl%') AND time >= datetime('now', '-{{lookback_days}} days') GROUP BY user_agent, actor_user_name, provider HAVING accounts <= 2What a hit looks like. User-Agents corresponding to scripting libraries or CLIs seen on very few cloud accounts. Normal administrative scripts might appear, but framework defaults will be distinct.
-
Cloud CLI and scripting on endpoints
Query · enrichmentCorrelate cloud API activity with local process execution of cloud CLI tools or scripting interpreters, scoped to the hosts from the first step.
reads hb_process_activitysqlSELECT device_hostname, process_name, process_cmd_line, user_name, time FROM hb_process_activity WHERE (LOWER(process_name) LIKE '%aws%' OR LOWER(process_name) LIKE '%az.exe%' OR LOWER(process_name) LIKE '%gcloud%' OR LOWER(process_name) LIKE '%python%' OR LOWER(process_name) LIKE '%terraform%') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND time >= datetime('now', '-{{lookback_days}} days')What a hit looks like. Execution of cloud management tools or scripting interpreters on the identified hosts. Silence means no such tools ran on the scoped hosts.
-
Triage Identity and API Evidence
Agent triageSynthesize sign-in data, user-agent rarity, and local process activity to determine if an agentic loop is active.
-
Evaluate Triage Verdict
DecisionRoute to containment if an agentic framework is confirmed.
-
Revoke Cloud Sessions
Response actionNeutralize the automated agent by revoking tokens and forcing a re-authentication.
-
Analyst review and tuning
Analyst taskVerify the agent's findings and distinguish malicious automation from legitimate administrative work.
-
Close out
Analyst taskDocument findings and update baseline for known automation.
Coverage
Scenario coverage
| Stage | Covered | How, or why not |
|---|---|---|
| Agentic Cloud and Identity Access T1078 |
Yes | untrusted-cloud-sign-ins, rare-automation-user-agents, endpoint-cloud-tooling |
| Initial Access via Extensions and Exploits T1176 · T1190 |
Out of scope | Belongs to another part of the 'Inside the loop: Five AI attack frameworks that never ask permission' series. |
| High-Tempo Command and Tool Execution T1059 · T1558.003 |
Out of scope | Belongs to another part of the 'Inside the loop: Five AI attack frameworks that never ask permission' series. |
| Persistence via Implants and WebShells T1505.003 · T1547.001 · T1053.005 |
Out of scope | Belongs to another part of the 'Inside the loop: Five AI attack frameworks that never ask permission' series. |
| C2 and Proxy Infrastructure T1090.003 |
Out of scope | Belongs to another part of the 'Inside the loop: Five AI attack frameworks that never ask permission' series. |
Blind spots
- Needs Detailed CloudTrail Data Events or UAL Advanced Audit. A hunt relying only on management events will miss high-tempo data exfiltration bursts. It would answer whether individual read/write operations inside a storage bucket were performed.
- Needs hb_process_activity with container context. If the endpoint agent is not present on the container host, the process activity will be invisible, leaving only the cloud API logs. It would answer whether the framework ran inside a short-lived container (e.g. Cairn).
Parameters & data
Parameters
| Parameter | Type | Default | What it is |
|---|---|---|---|
lookback_days | number | 14 | Days of history to examine. |
scope_hosts | list[host] | — | Workstation or server names identified in the sign-in scoping to focus endpoint process analysis. |
target_providers | list[string] | aws, github, azure_ad, m365, slack | Cloud and SaaS providers targeted by agentic frameworks. |
Telemetry
| Source | Category | Telemetry |
|---|---|---|
| Endpoint telemetry (hb_ surfaces) | endpoint | endpoint |
| Identity / sign-in telemetry | identity | identity |
Source
---
analysis: A single detection rule might flag a suspicious User-Agent, but it cannot
correlate the 'Identity + Automation + Endpoint' triad. The hunt context allows
an analyst to weigh whether a rare Boto3 agent is a new developer or an ARTEX worker.
blind_spots:
- id: limited-cloud-logging
question: whether individual read/write operations inside a storage bucket were
performed
requires: Detailed CloudTrail Data Events or UAL Advanced Audit
risk: A hunt relying only on management events will miss high-tempo data exfiltration
bursts.
stage: agentic-cloud-identity-access
- id: ephemeral-infrastructure
question: whether the framework ran inside a short-lived container (e.g. Cairn)
requires: hb_process_activity with container context
risk: If the endpoint agent is not present on the container host, the process activity
will be invisible, leaving only the cloud API logs.
stage: agentic-cloud-identity-access
coverage:
- stage: agentic-cloud-identity-access
status: covered
steps:
- untrusted-cloud-sign-ins
- rare-automation-user-agents
- endpoint-cloud-tooling
- reason: 'Belongs to another part of the ''Inside the loop: Five AI attack frameworks
that never ask permission'' series.'
stage: initial-access-extensions-and-exploits
status: out_of_scope
- reason: 'Belongs to another part of the ''Inside the loop: Five AI attack frameworks
that never ask permission'' series.'
stage: high-tempo-command-execution
status: out_of_scope
- reason: 'Belongs to another part of the ''Inside the loop: Five AI attack frameworks
that never ask permission'' series.'
stage: persistence-via-implants-and-webshells
status: out_of_scope
- reason: 'Belongs to another part of the ''Inside the loop: Five AI attack frameworks
that never ask permission'' series.'
stage: c2-and-proxy-infrastructure
status: out_of_scope
guardrails:
claims: no_unsupported
evidence: citation_required
missing_data: not_benign
telemetry: untrusted
hunt:
applicability: campaign-specific
handoff: keep-as-periodic-hunt
justification: Agentic frameworks operate at a tempo that defeats manual response
and single-behavior logic; a coordinated hunt for the 'unmanaged-access + automated-burst'
pattern is required to identify automated persistence in cloud control planes.
methodology: model-assisted
trigger: intel-report
hypothesis: An intruder is using an agentic framework to automate cloud API interactions
and identity access, characterized by successful sign-ins from unmanaged devices
followed by high-tempo API bursts and local CLI tool usage.
labels:
- hunt
- attack.t1078
- attack.t1059
- attack.t1090.003
- attack.t1505.003
- command and control
- execution
- initial access
- persistence
name: Agentic Identity and Cloud Control Plane Exploitation
parameters:
lookback_days:
default: '14'
description: Days of history to examine.
type: number
scope_hosts:
default: []
description: Workstation or server names identified in the sign-in scoping to
focus endpoint process analysis.
type: list[host]
target_providers:
default:
- aws
- github
- azure_ad
- m365
- slack
description: Cloud and SaaS providers targeted by agentic frameworks.
from:
kind: article
observed: '2026-11-13'
ref: https://www.elastic.co/security-labs/threat-command/agentic-offensive-frameworks-detection-artex
type: list[string]
provenance:
authors:
- name: Huntbase hunt generation
org: huntbase.io
generated:
by: huntbase-hunt-generation
from: https://www.elastic.co/security-labs/threat-command/agentic-offensive-frameworks-detection-artex
gates:
- dry-run
- lint
- critic
model: hb_google/gemini-3-flash-preview
rationale: Focus on AWS, Entra ID, and GitHub logs. Look for successful sign-ins where
device compliance is false or missing, then use those source hosts to narrow the
endpoint process search.
references:
- name: 'Inside the loop: Five AI attack frameworks that never ask permission'
url: https://www.elastic.co/security-labs/threat-command/agentic-offensive-frameworks-detection-artex
related:
- hunt: high-tempo-command-execution
reason: This hunt identifies the cloud entry, while the follow-on hunt looks at
the host-level command bursts.
relation: follows
scenario:
stages:
- name: Initial Access via Extensions and Exploits
observables:
- ARTEX browser extension recording authenticated sessions
- Vulnerability discovery scanning using stock settings
- Exploitation of public-facing web applications
- User-Agent strings identifying frameworks
slug: initial-access-extensions-and-exploits
tactic: initial-access
techniques:
- T1176
- T1190
- name: Agentic Cloud and Identity Access
observables:
- Sign-ins to AWS, Entra ID, Slack, and GitHub
- Abuse of recorded session tokens from extensions
- Automation of cloud CLI tools for account access
slug: agentic-cloud-identity-access
tactic: initial-access
techniques:
- T1078
- name: High-Tempo Command and Tool Execution
observables:
- Command bursts occurring in 2-25 second intervals
- Execution of curl, shells, and cloud CLIs
- Model Context Protocol (MCP) server function calls (HexStrike)
- Reflector agents sending human-like message responses (PentAGI)
- Credential access actions including Kerberoasting
slug: high-tempo-command-execution
tactic: execution
techniques:
- T1059
- T1558.003
- name: Persistence via Implants and WebShells
observables:
- CyberStrikeAI WebShell manager connections
- Creation of custom implants checking in every 5 seconds
- Registry Run keys for implant startup
- Windows Scheduled Tasks for recurring execution
- Dropping of coding agent workers in containers (Cairn)
slug: persistence-via-implants-and-webshells
tactic: persistence
techniques:
- T1505.003
- T1547.001
- T1053.005
- name: C2 and Proxy Infrastructure
observables:
- Connections to Telegram, Slack, Discord, and DingTalk C2 listeners
- DNS queries for update.pentagi.com and com.apple.systemupdate
- Traffic targeting 224.0.0.0
- Routing through multi-hop proxies to bypass IP bans
slug: c2-and-proxy-infrastructure
tactic: command-and-control
techniques:
- T1090.003
summary: Agentic offensive frameworks like ARTEX and CyberStrikeAI automate attack
sequences using LLM planners to orchestrate tools like scanners, cloud CLIs, and
custom implants in high-tempo bursts. These frameworks frequently bypass human
authorization, interacting with cloud control planes and host endpoints via rapid-fire
command execution and persistent backdoors.
series:
index: 1
slug: inside-the-loop-five-ai-attack-frameworks-that-never-ask-permission
title: 'Inside the loop: Five AI attack frameworks that never ask permission'
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
---
# Agentic Identity and Cloud Control Plane Exploitation
This hunt identifies the footprint of agentic offensive frameworks like ARTEX and CyberStrikeAI in cloud environments. These frameworks automate the 'think-time' of an operator, issuing API commands in bursts of 2 to 25 seconds. We look for successful authentications to cloud and SaaS providers (AWS, GitHub, Entra ID) from unmanaged devices or unknown hosts, then correlate these identities with high-frequency API activity and the use of automated User-Agents or local cloud CLI tools that are rare across the environment. An agent evaluates the temporal proximity of the sign-in to the API bursts to distinguish automated agents from human administrators.
## untrusted-cloud-sign-ins
<!-- Cloud sign-ins from unmanaged devices -->
Identify successful authentications to critical cloud platforms originating from unmanaged devices or unknown workstations, providing the lead for session abuse.
```sqlite target=identity role=scoping params=(lookback_days=lookback_days, target_providers=target_providers)
~~~yaml
expected: A list of users and source hosts that signed into cloud services without
a managed device context. Silence suggests all sign-ins were from trusted devices.
reads:
- actor_user_name
- src_endpoint_ip
- src_endpoint_hostname
- device_hostname
- provider
- device_trust_type
- device_is_compliant
- status_id
- activity_id
- time
silence: not_evidence_of_absence
source: hb_auth_signin
verified: dry-run
verified_at: '2026-11-14'
~~~
SELECT actor_user_name, src_endpoint_ip, src_endpoint_hostname, device_hostname, provider, device_trust_type, time FROM hb_auth_signin WHERE activity_id = 1 AND status_id = 1 AND (device_is_compliant = 'false' OR device_trust_type IS NULL OR device_trust_type = '') AND instr(',' || '{{target_providers}}' || ',', ',' || LOWER(provider) || ',') > 0 AND time >= datetime('now', '-{{lookback_days}} days')
```
## analyze-activity
<!-- Analyze behavior for automation and bursts -->
parallel:
- → rare-automation-user-agents
- → endpoint-cloud-tooling
join: → triage-agentic-activity
## rare-automation-user-agents
<!-- Rare automated User-Agents in cloud API logs -->
Detect non-browser tools like Boto3 or Python-requests that are rare across the environment, which often facilitate agentic framework interactions.
```sqlite target=endpoint role=baseline params=(lookback_days=lookback_days)
~~~yaml
baseline:
compare: first_seen
window: '{{lookback_days}}d'
expected: User-Agents corresponding to scripting libraries or CLIs seen on very few
cloud accounts. Normal administrative scripts might appear, but framework defaults
will be distinct.
prevalence:
by: cloud_account_uid
key:
- user_agent
rare_below: 3
reads:
- user_agent
- actor_user_name
- provider
- cloud_account_uid
- time
silence: not_evidence_of_absence
source: hb_cloud_api_activity
verified: dry-run
verified_at: '2026-11-14'
~~~
SELECT user_agent, actor_user_name, provider, COUNT(DISTINCT cloud_account_uid) AS accounts, COUNT(*) AS calls, MIN(time) AS first_seen FROM hb_cloud_api_activity WHERE (LOWER(user_agent) LIKE '%python%' OR LOWER(user_agent) LIKE '%boto%' OR LOWER(user_agent) LIKE '%aws-cli%' OR LOWER(user_agent) LIKE '%curl%') AND time >= datetime('now', '-{{lookback_days}} days') GROUP BY user_agent, actor_user_name, provider HAVING accounts <= 2
```
## endpoint-cloud-tooling
<!-- Cloud CLI and scripting on endpoints -->
Correlate cloud API activity with local process execution of cloud CLI tools or scripting interpreters, scoped to the hosts from the first step.
```sqlite target=endpoint role=enrichment params=(lookback_days=lookback_days, scope_hosts=scope_hosts)
~~~yaml
expected: Execution of cloud management tools or scripting interpreters on the identified
hosts. Silence means no such tools ran on the scoped hosts.
reads:
- device_hostname
- process_name
- process_cmd_line
- user_name
- time
silence: not_evidence_of_absence
source: hb_process_activity
verified: dry-run
verified_at: '2026-11-14'
~~~
SELECT device_hostname, process_name, process_cmd_line, user_name, time FROM hb_process_activity WHERE (LOWER(process_name) LIKE '%aws%' OR LOWER(process_name) LIKE '%az.exe%' OR LOWER(process_name) LIKE '%gcloud%' OR LOWER(process_name) LIKE '%python%' OR LOWER(process_name) LIKE '%terraform%') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND time >= datetime('now', '-{{lookback_days}} days')
```
## triage-agentic-activity
<!-- Triage Identity and API Evidence -->
```agent target=hunter
cite: required
context:
- untrusted-cloud-sign-ins
- rare-automation-user-agents
- endpoint-cloud-tooling
max_iterations: 5
objective: 'Determine if the combination of unmanaged cloud sign-ins, rare automation
user-agents, and local cloud CLI usage indicates an automated offensive framework
like ARTEX or CyberStrikeAI. Analyze the timing: do the API calls and local processes
start immediately after the unmanaged sign-in?'
success_criteria: A verdict of malicious | suspicious | benign per identity, citing
specific rows and timestamps.
tools:
- endpoint
- identity
```
## evaluate-verdict
<!-- Evaluate Triage Verdict -->
if~: "the triage verdict is malicious for at least one user identity" (confidence: high, judge=hunter)
then: → revoke-sessions
indeterminate: → analyst-manual-review
unavailable: → analyst-manual-review (blind_spot: limited-cloud-logging)
else: → close-out
## revoke-sessions
<!-- Revoke Cloud Sessions -->
```action target=identity
~~~yaml
approval: required
~~~
Revoke all active sessions and refresh tokens for the identified user across the impacted cloud providers. Reset credentials and rotate access keys if applicable.
```
→ analyst-manual-review
## analyst-manual-review
<!-- Analyst review and tuning -->
```manual target=analyst
Review the identified User-Agents and local process command lines. Correlate with internal engineering schedules to ensure these weren't legitimate automation tasks. If malicious, initiate the full incident response process.
```
→ close-out
## close-out
<!-- Close out -->
```manual target=analyst
Document the identified automated User-Agents. If benign, add them to a known-automation allow-list for future periodic hunts.
```
→ 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, critic, then reviewed by a person.