Skip to content
.ca
4 mininfo

Building an Adversarial Consensus Engine | Multi-Agent LLMs for Automated Malware Analysis

SentinelOne Labs developed a multi-agent LLM architecture using OpenClaw and Claude models to automate malware reverse engineering. By employing a serial consensus pipeline with an active rejection mandate, the system forces independent tool agents (radare2, Ghidra, Binary Ninja, IDA Pro) to cross-validate findings, significantly reducing decompiler artifacts and hallucinations.

Conf:highAnalyzed:2026-04-23reports

Authors: SentinelOne Labs

ActorsSysJokerWizardUpdateGo InfostealerFinderRAT

Source:SentinelOne

IOCs · 4

Key Takeaways

  • Multi-agent LLM architectures using serial consensus pipelines significantly reduce hallucinations in automated malware analysis.
  • An 'Active Rejection Mandate' forces LLM agents to peer-review and explicitly reject decompiler artifacts from other tools.
  • Deterministic bridge scripts are preferred over Model Context Protocol (MCP) for batch analysis due to lower latency and comprehensive extraction.
  • Testing on macOS malware demonstrated the system's ability to cross-validate XOR-obfuscated payloads, hidden execution paths, and C2 endpoints.

Affected Systems

  • macOS

Attack Chain

The analyzed macOS malware utilizes XOR-obfuscated payloads to hide its core capabilities from static analysis. Upon execution, it performs system fingerprinting by collecting the username, hardware UUID, and macOS version. It establishes persistence via a LaunchAgent plist configured to run hourly. Finally, it beacons to a remote C2 server using curl to exfiltrate the collected system data and potentially download further payloads via ditto archive extraction.

Detection Availability

  • YARA Rules: No
  • Sigma Rules: No
  • Snort/Suricata Rules: No
  • KQL Queries: No
  • Splunk SPL Queries: No
  • EQL Queries: No
  • Other Detection Logic: No

The article focuses on automated malware analysis methodology and does not provide specific detection rules.

Detection Engineering Assessment

EDR Visibility: High — EDRs on macOS can effectively monitor process executions like curl and ditto, as well as file modifications in LaunchAgents directories. Network Visibility: Medium — C2 traffic is HTTPS (curl POST), limiting payload inspection without SSL decryption, but domain and IP connections are visible. Detection Difficulty: Moderate — While the payload strings are XOR obfuscated, the behavioral indicators (LaunchAgent creation, curl POSTs, ditto extraction) are standard and detectable.

Required Log Sources

  • Process Execution Logs
  • File Creation Logs
  • Network Connection Logs

Hunting Hypotheses

HypothesisTelemetryATT&CK StageFP Risk
Look for curl executions making POST requests with JSON payloads containing system fingerprinting data (event, machine, url, os_name).Process Execution, Network ConnectionsExfiltrationLow
Monitor for the creation of LaunchAgent plist files with RunAtLoad set to true and a StartInterval of 3600.File CreationPersistenceMedium
Hunt for processes querying the QuarantineEventsV2 database to extract download URLs.File Access, Process ExecutionDiscoveryLow
Detect the use of 'ditto -x -k' for silent archive extraction in suspicious contexts or by unexpected parent processes.Process ExecutionExecutionMedium

Control Gaps

  • Static analysis tools may miss capabilities hidden behind XOR obfuscation without dynamic execution or advanced multi-tool consensus.

Key Behavioral Indicators

  • ditto -x -k execution
  • curl POST to unknown domains
  • LaunchAgent creation with specific intervals

False Positive Assessment

  • Low

Recommendations

Immediate Mitigation

  • Block the identified C2 domain (eve.inspectorcleaner.com) at the network perimeter.
  • Hunt for the provided SHA256 hashes in the environment to identify potential compromises.

Infrastructure Hardening

  • Implement strict egress filtering to prevent unauthorized curl communications to unknown external domains.

User Protection

  • Ensure macOS endpoints are running modern EDR solutions capable of monitoring LaunchAgent persistence mechanisms.

Security Awareness

  • N/A

MITRE ATT&CK Mapping

  • T1027 - Obfuscated Files or Information
  • T1140 - Deobfuscate/Decode Files or Information
  • T1082 - System Information Discovery
  • T1543.001 - Create or Modify System Process: Launch Agent
  • T1059.004 - Command and Scripting Interpreter: Unix Shell
  • T1071.001 - Application Layer Protocol: Web Protocols

Additional IOCs

  • Domains:
    • eve[.]inspectorcleaner[.]com - C2 domain extracted from malware payload.
  • Urls:
    • hxxps://eve[.]inspectorcleaner[.]com/vnt - C2 beacon endpoint.
  • File Hashes:
    • 60c8128c48aac890a6d01448d1829a6edcdce0d2 (SHA256) - WizardUpdate sample
    • 678aa572faa73f6873d24f24e423d315e7eb2c2d (SHA256) - Go Infostealer sample
    • ad7d2eb98ea4ddc7700db786aadb796b286da04 (SHA256) - FinderRAT sample
    • f5149543014e5b1bd7030711fd5c7d2a4bef0c2f (SHA256) - SysJoker sample
  • File Paths:
    • /bin/bash - Hidden execution path utilized by the malware.
    • /api/req/res - SysJoker C2 API endpoint.
  • Command Lines:
    • Purpose: C2 beaconing and data exfiltration | Tools: curl | Stage: Exfiltration | curl POST
    • Purpose: Archive extraction for payload delivery | Tools: ditto | Stage: Execution | ditto -x -k
  • Other:
    • QuarantineEventsV2 - macOS Quarantine database queried by the malware to find its own download URL.
    • OIUSYTASVBNMOPASDCCVBNZAXX - XOR key used for string obfuscation.