Connection to Common Large Language Model Endpoints

IMPORTANT: This documentation is no longer updated. Refer to Elastic's version policy and the latest documentation.

Connection to Common Large Language Model Endpoints

Identifies DNS queries to known Large Language Model domains by unsigned binaries or common Windows scripting utilities. Malwares may leverage the capabilities of LLM to perform actions in the affected system in a dynamic way.

Rule type: eql

Rule indices:

endgame-*
logs-endpoint.events.network-*
logs-sentinel_one_cloud_funnel.*
logs-windows.sysmon_operational-*
winlogbeat-*

Severity: medium

Risk score: 47

Runs every: 5m

Searches indices from: now-9m (Date Math format, see also Additional look-back time)

Maximum alerts per execution: 100

References:

https://malpedia.caad.fkie.fraunhofer.de/details/py.lamehug

Tags:

Domain: Endpoint
OS: Windows
Use Case: Threat Detection
Tactic: Command and Control
Resources: Investigation Guide
Data Source: Elastic Endgame
Data Source: Elastic Defend
Data Source: SentinelOne
Data Source: Sysmon

Version: 1

Rule authors:

Elastic

Rule license: Elastic License v2

Investigation guide

edit

Triage and analysis

Investigating Connection to Common Large Language Model Endpoints

Possible investigation steps

Investigate the process execution chain (parent process tree) for unknown processes or malicious scripts.
Verify if the executed process is persistent on the host like common mechanisms Startup folder, task or Run key.
Review any unusual network, files or registry events by the same process.
Investigate other alerts associated with the user/host during the past 48 hours.
Extract this communication’s indicators of compromise (IoCs) and use traffic logs to search for other potentially compromised hosts.

False positive analysis

Trusted applications from an expected process running in the environment.

Response and remediation

Initiate the incident response process based on the outcome of the triage.
Isolate the involved host to prevent further post-compromise behavior.
Immediately block the identified indicators of compromise (IoCs).
Implement any temporary network rules, procedures, and segmentation required to contain the attack.
Investigate credential exposure on systems compromised or used by the attacker to ensure all compromised accounts are identified. Reset passwords for these accounts and other potentially compromised credentials, such as email, business systems, and web services.
Update firewall rules to be more restrictive.
Reimage the host operating system or restore the compromised files to clean versions.
Run a full antimalware scan. This may reveal additional artifacts left in the system, persistence mechanisms, and malware components.
Determine the initial vector abused by the attacker and take action to prevent reinfection through the same vector.
Using the incident response data, update logging and audit policies to improve the mean time to detect (MTTD) and the mean time to respond (MTTR).

Rule query

edit

network where host.os.type == "windows" and dns.question.name != null and
(
  process.name : ("MSBuild.exe", "mshta.exe", "wscript.exe", "powershell.exe", "pwsh.exe", "msiexec.exe", "rundll32.exe",
  "bitsadmin.exe", "InstallUtil.exe", "RegAsm.exe", "vbc.exe", "RegSvcs.exe", "python.exe", "regsvr32.exe", "dllhost.exe",
  "node.exe", "javaw.exe", "java.exe", "*.pif", "*.com") or

  ?process.code_signature.subject_name : ("AutoIt Consulting Ltd", "OpenJS Foundation", "Python Software Foundation") or

  (process.executable : ("?:\\Users\\*.exe", "", "?:\\ProgramData\\*.exe") and ?process.code_signature.trusted != true)
 ) and
    dns.question.name : (
    // Major LLM APIs
    "api.openai.com",
    "*.openai.azure.com",
    "api.anthropic.com",
    "api.mistral.ai",
    "api.cohere.ai",
    "api.ai21.com",
    "api.groq.com",
    "api.perplexity.ai",
    "api.x.ai",
    "api.deepseek.com",
    "api.gemini.google.com",
    "generativelanguage.googleapis.com",
    "api.azure.com",
    "api.bedrock.aws",
    "bedrock-runtime.amazonaws.com",

    // Hugging Face & other ML infra
    "api-inference.huggingface.co",
    "inference-endpoint.huggingface.cloud",
    "*.hf.space",
    "*.replicate.com",
    "api.replicate.com",
    "api.runpod.ai",
    "*.runpod.io",
    "api.modal.com",
    "*.forefront.ai",

    // Consumer-facing AI chat portals
    "chat.openai.com",
    "chatgpt.com",
    "copilot.microsoft.com",
    "bard.google.com",
    "gemini.google.com",
    "claude.ai",
    "perplexity.ai",
    "poe.com",
    "chat.forefront.ai",
    "chat.deepseek.com"
  ) and

  not process.executable : (
          "?:\\Program Files\\*.exe",
          "?:\\Program Files (x86)\\*.exe",
          "?:\\Windows\\System32\\svchost.exe",
          "?:\\Windows\\SystemApps\\Microsoft.LockApp_*\\LockApp.exe",
          "?:\\Users\\*\\AppData\\Local\\Google\\Chrome\\Application\\chrome.exe",
          "?:\\Users\\*\\AppData\\Local\\BraveSoftware\\*\\Application\\brave.exe",
          "?:\\Users\\*\\AppData\\Local\\Vivaldi\\Application\\vivaldi.exe",
          "?:\\Users\\*\\AppData\\Local\\Programs\\Opera*\\opera.exe",
          "?:\\Users\\*\\AppData\\Local\\Programs\\Fiddler\\Fiddler.exe"
        ) and
    not (?process.code_signature.trusted == true and
         ?process.code_signature.subject_name : ("Anthropic, PBC", "Google LLC", "Mozilla Corporation", "Brave Software, Inc.", "Island Technology Inc.", "Opera Norway AS"))

Framework: MITRE ATT&CK^TM

Tactic:
- Name: Command and Control
- ID: TA0011
- Reference URL: https://attack.mitre.org/tactics/TA0011/
Technique:
- Name: Web Service
- ID: T1102
- Reference URL: https://attack.mitre.org/techniques/T1102/

« Multi-Base64 Decoding Attempt from Suspicious Location Network Activity to a Suspicious Top Level Domain »