Threat actors are weaponizing agentic AI frameworks to scale attacks faster than traditional human-operated campaigns. Amazon's threat intelligence team recently identified criminal groups abusing ARTEX — an open-source agentic AI penetration testing framework — to automate reconnaissance, credential harvesting, and deepfake-based identity bypass attempts.
Between October 3 and November 5, 2025, Amazon's MadPot honeypot network observed ARTEX agents probing decoy sensors, attempting to steal AI credentials, and testing deepfake ID verification bypasses. This marks a concerning shift: adversaries are now deploying autonomous AI agents that can execute multi-stage attacks with minimal human oversight.
For defenders, this isn't theoretical. AI-driven attack agents can enumerate targets, test credentials, adapt evasion techniques, and operate at machine speed — overwhelming traditional SOC workflows. Organizations must adapt their detection engineering and hardening strategies to counter AI-automated threats now.
Technical Analysis
What is ARTEX?
ARTEX (Automated Red Team EXecution) is an open-source penetration testing framework designed to simulate adversarial TTPs using AI agents. While intended for legitimate security testing, threat actors have weaponized it to:
- Steal AI credentials: ARTEX agents target exposed API keys, LLM service tokens (e.g., OpenAI, Anthropic, AWS Bedrock), and cloud AI service credentials.
- Test deepfake ID bypasses: Criminals use ARTEX to automate deepfake video/audio generation and test bypasses against identity verification systems (KYC/AML, biometric authentication).
- Automate reconnaissance: ARTEX can scan for misconfigured cloud AI endpoints, enumerate exposed LLM services, and fingerprint authentication mechanisms.
Attack Chain
- Reconnaissance: ARTEX agents scan the internet for exposed AI services, cloud APIs, and identity verification endpoints.
- Credential Harvesting: Automated probing for leaked or weakly protected AI API keys, service account tokens, and authentication credentials.
- Deepfake Testing: ARTEX generates synthetic media and tests it against ID verification systems to identify bypass opportunities.
- Exploitation: Stolen credentials are used to access LLM services, cloud AI platforms, or downstream systems for fraud, espionage, or further compromise.
Affected Platforms
- Cloud AI Services: AWS Bedrock, Azure OpenAI, Google Vertex AI, Anthropic, OpenAI API
- Identity Verification Systems: KYC/AML platforms, biometric authentication services, video-based ID verification
- CI/CD Pipelines: Environments with hardcoded AI credentials in code repositories or environment variables
Exploitation Status
- In-the-Wild: Confirmed active abuse by criminal groups (as reported by Amazon).
- No CVE: ARTEX itself is not a vulnerability — it's a tool being misused. The risk comes from exposed credentials and weak identity verification controls.
- CISA KEV: Not applicable (no specific CVE).
Detection & Response
SIGMA Rules
---
title: AI Credential Access via ARTEX Agent
description: Detects suspicious process or script attempting to access cloud AI service credentials (AWS Bedrock, OpenAI, Azure OpenAI) from unusual locations or with automation patterns.
status: experimental
author: Security Arsenal
references:
- https://attack.mitre.org/techniques/T1552/001/
- https://attack.mitre.org/techniques/T1078/004/
logsource:
category: process_creation
product: windows
detection:
selection:
Image|endswith:
- '\python.exe'
- '\pythonw.exe'
- '\powershell.exe'
- '\pwsh.exe'
CommandLine|contains:
- 'bedrock'
- 'openai'
- 'azure-ai'
- 'anthropic'
- 'vertex-ai'
filter:
Image|startswith: 'C:\Program Files\'
condition: selection and not filter
level: medium
tags:
- attack.credential_access
- attack.t1552.001
---
title: Deepfake Video Generation Tool Execution
description: Detects execution of deepfake generation tools (DeepFaceLab, FaceSwap, etc.) on non-designated systems.
status: experimental
author: Security Arsenal
references:
- https://attack.mitre.org/techniques/T1204/002/
logsource:
category: process_creation
product: windows
detection:
selection:
Image|endswith:
- '\DeepFaceLab.exe'
- '\faceswap.exe'
- '\deepfake.exe'
CommandLine|contains:
- 'deepfake'
- 'faceswap'
- 'deepfacelab'
filter:
Image|startswith: 'C:\Users\'
condition: selection and not filter
level: high
tags:
- attack.impact
- attack.t1204.002
---
title: Suspicious Network Connections to AI Service Endpoints
description: Detects outbound connections to cloud AI service APIs from non-browser processes or unusual hosts.
status: experimental
author: Security Arsenal
references:
- https://attack.mitre.org/techniques/T1071/001/
logsource:
category: network_connection
product: windows
detection:
selection:
DestinationHost|contains:
- 'bedrock.amazonaws.com'
- 'openai.azure.com'
- 'api.openai.com'
- 'vertexai.googleapis.com'
- 'api.anthropic.com'
filter:
Image|endswith:
- '\chrome.exe'
- '\firefox.exe'
- '\msedge.exe'
- '\iexplore.exe'
condition: selection and not filter
level: medium
tags:
- attack.command_and_control
- attack.t1071.001
KQL (Microsoft Sentinel / Defender)
// Hunt for ARTEX-like activity: processes accessing AI credentials or deepfake tools
let suspiciousKeywords = dynamic(['bedrock', 'openai', 'azure-ai', 'anthropic', 'vertex-ai', 'deepfake', 'faceswap', 'deepfacelab']);
union withsource=tableName *
| where TimeGenerated > ago(7d)
| where tableName == "SecurityEvent" or tableName == "DeviceProcessEvents"
| extend ProcessCommandLine = tolower(tostring(parse_json(AdditionalFields).CommandLine))
| where ProcessCommandLine has_any (suspiciousKeywords)
| where not (ProcessCommandLine contains "program files")
| project TimeGenerated, Computer, Account, Process, ProcessCommandLine, tableName
| order by TimeGenerated desc
Velociraptor VQL
-- Hunt for processes accessing AI service credentials or deepfake tools
SELECT Pid, Name, Exe, CommandLine, Username, CreateTime
FROM pslist()
WHERE CommandLine =~ '(bedrock|openai|azure-ai|anthropic|vertex-ai|deepfake|faceswap|deepfacelab)'
AND NOT CommandLine =~ '(program files|system32)'
Remediation Script (PowerShell)
# ARTEX Defense: Harden AI credential storage and detect deepfake tools
# Run as Administrator
# 1. Audit for hardcoded AI credentials in environment variables
Write-Host "[+] Scanning for AI credentials in environment variables..."
$envVars = Get-ChildItem Env:
foreach ($var in $envVars) {
if ($var.Value -match '(sk-|api_key|token|secret)' -and $var.Name -match '(openai|bedrock|azure|anthropic|vertex)') {
Write-Warning "Potential AI credential found in environment variable: $($var.Name)"
}
}
# 2. Block execution of known deepfake tools via AppLocker (example policy)
Write-Host "[+] Creating AppLocker policy to block deepfake tools..."
$deepfakeTools = @("DeepFaceLab.exe", "faceswap.exe", "deepfake.exe")
foreach ($tool in $deepfakeTools) {
$rule = New-AppLockerPolicy -RuleType Publisher -User "Everyone" -Action Deny -Publisher "*" -ProductName "*" -FileName $tool
Set-AppLockerPolicy -XmlPolicy $rule.ToXml() -Merge
}
Write-Host "AppLocker policy updated. Deepfake tools are now blocked for all users."
# 3. Enable advanced audit policies for credential access
Write-Host "[+] Enabling advanced audit policies..."
auditpol /set /subcategory:"Sensitive Privilege Use" /success:enable /failure:enable
auditpol /set /subcategory:"Process Creation" /success:enable /failure:enable
Write-Host "[+] Hardening complete. Monitor for alerts on AI credential access and deepfake tool execution."
Remediation
Immediate Actions
- Rotate AI Credentials: Immediately rotate any exposed or hardcoded API keys for cloud AI services (AWS Bedrock, OpenAI, Azure OpenAI, etc.).
- Audit Code Repositories: Scan for hardcoded AI credentials in source code, CI/CD pipelines, and configuration files.
- Harden Identity Verification: If you operate KYC/AML or biometric systems, review deepfake detection capabilities and update to the latest anti-spoofing models.
- Block Deepfake Tools: Use application whitelisting (AppLocker, WDAC) to block known deepfake generation tools on non-designated systems.
Long-Term Defenses
- Implement AI Credential Vaulting: Use secrets management solutions (HashiCorp Vault, AWS Secrets Manager, Azure Key Vault) to store AI service credentials — never hardcode them.
- Monitor AI Service Usage: Deploy cloud-native monitoring (AWS CloudTrail, Azure Monitor) to detect unusual AI API calls or credential access patterns.
- Adopt AI-Powered Detection: Use behavioral analytics and ML-based SIEM rules to detect anomalous AI service access (e.g., unusual API call volumes, geographic anomalies).
- Update Deepfake Detection: Regularly update identity verification systems with the latest deepfake detection models and liveness checks.
- Threat Intelligence: Subscribe to threat feeds for AI-related attack patterns and ARTEX framework indicators.
Vendor Resources
- AWS: Bedrock Security Best Practices
- Microsoft: Azure OpenAI Security
- CISA: AI Risk Management
Related Resources
Security Arsenal Penetration Testing Services AlertMonitor Platform Book a SOC Assessment vulnerability-management Intel Hub
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.