AI as tradecraft: How threat actors operationalize AI
Threat actors, particularly North Korean state-sponsored groups, are increasingly operationalizing AI to accelerate cyberattacks. They leverage generative AI for reconnaissance, social engineering, identity fabrication, and malware development, acting as a force multiplier that reduces technical friction while human operators maintain control over objectives.
Authors: Microsoft Threat Intelligence
Source:
Microsoft
- urlhxxp://144[.]172[.]105[.]122:8085/api/file/1/API endpoint used for file uploads in AI-assisted malware.
Key Takeaways
- Threat actors use AI as a force multiplier across the attack lifecycle, accelerating reconnaissance, resource development, and malware generation.
- North Korean actors (e.g., Jasper Sleet, Coral Sleet) heavily leverage AI for identity fabrication, deepfakes, and social engineering to secure remote IT jobs.
- AI is actively used to accelerate malware development, leaving distinct artifacts like emojis and conversational comments in the code.
- Emerging trends include threat actor experimentation with agentic AI workflows and AI recommendation poisoning.
- Defenders must shift from static indicators to behavioral signals to detect AI-generated phishing, domains, and malware.
Affected Systems
- Enterprise AI Systems
- Microsoft 365
- Identity Infrastructure
- Remote Hiring Platforms
Vulnerabilities (CVEs)
- CVE-2022-30190
Attack Chain
Threat actors begin by using LLMs for reconnaissance, researching vulnerabilities and target personas. They then use generative AI to develop resources, such as GAN-based look-alike domains and highly convincing fake identities with deepfake images and voice modulation. Initial access is achieved through AI-polished phishing lures or by getting hired as remote IT workers using fabricated credentials. Post-compromise, AI assists in discovering assets, refining privilege escalation scripts, and summarizing exfiltrated data for extortion.
Detection Availability
- YARA Rules: No
- Sigma Rules: No
- Snort/Suricata Rules: No
- KQL Queries: Yes
- Splunk SPL Queries: No
- EQL Queries: No
- Other Detection Logic: No
- Platforms: Microsoft Defender XDR, Microsoft Sentinel
Microsoft provides KQL queries for Microsoft Defender XDR to detect potentially spoofed emails and surface suspicious sign-in attempts associated with remote IT worker fraud.
Detection Engineering Assessment
EDR Visibility: Medium — AI-assisted malware and scripts may evade static signatures, but behavioral EDR can catch the resulting execution, lateral movement, and C2 connections. Network Visibility: Medium — Network telemetry can identify connections to known malicious infrastructure or anomalous data exfiltration, though AI-generated domains might bypass static reputation filters. Detection Difficulty: Hard — AI-generated content (phishing, code, domains) is designed to blend in with legitimate traffic and bypass traditional pattern-matching defenses.
Required Log Sources
- EmailLogs
- EntraIdSignInEvents
- Network connections
- Process execution logs
Hunting Hypotheses
| Hypothesis | Telemetry | ATT&CK Stage | FP Risk |
|---|---|---|---|
| Look for unusual sign-in events from unmanaged devices with atypical travel or impossible travel alerts, indicating potential remote IT worker fraud. | Identity/Entra ID logs | Initial Access | Medium |
| Search for inbound emails failing SPF/DKIM/DMARC checks but originating from domains visually similar to trusted partners, indicating AI-generated spoofing. | Email Gateway logs | Initial Access | Low |
| Analyze custom scripts or binaries for unusual conversational comments or emoji-based logging, which may indicate AI-assisted malware generation. | File/Script analysis | Execution | High |
Control Gaps
- Static pattern-based phishing detection
- Traditional domain reputation filtering
- Identity verification in remote hiring
Key Behavioral Indicators
- Conversational in-line comments in scripts
- Emojis used as execution state markers in code
- Overly descriptive or redundant variable naming
- Over-engineered modular structure in simple scripts
False Positive Assessment
- Medium (Detecting AI-generated code or text based on style or emojis can lead to false positives, as legitimate developers also use AI assistants and emojis in their workflows.)
Recommendations
Immediate Mitigation
- Enforce MFA on all accounts and require it from all devices and locations.
- Turn on Zero-hour auto purge (ZAP) in Defender for Office 365 to retroactively neutralize malicious messages.
Infrastructure Hardening
- Configure Safe Links policies for internal recipients.
- Enable network protection in Microsoft Defender for Endpoint.
- Implement Azure AI Content Safety Prompt Shields to protect enterprise AI deployments from prompt injection.
User Protection
- Treat fraudulent employment as an insider-risk scenario and monitor for abnormal access patterns.
- Use Purview Insider Risk Management to detect data leakage and unauthorized access.
Security Awareness
- Conduct attack simulation training with realistic AI-generated phishing scenarios.
- Train HR and hiring managers to identify deepfakes and AI-generated personas during interviews.
MITRE ATT&CK Mapping
- T1589 - Gather Victim Identity Information
- T1583 - Acquire Infrastructure
- T1566 - Phishing
- T1078 - Valid Accounts
- T1059 - Command and Scripting Interpreter
Additional IOCs
- Ips:
144[.]172[.]105[.]122- C2 or exfiltration server IP identified in an AI-generated malware code snippet.
- Urls:
hxxp://144[.]172[.]105[.]122:8085- Base URL for C2 server found in AI-generated script.
- Other:
✅- Visual marker emoji used in AI-generated code paths for successful execution.❌- Visual marker emoji used in AI-generated code paths for indicating errors.