AI-Powered / Autonomous Malware

Introduction

AI-powered malware hypothetically uses machine learning to autonomously adapt its behaviour in response to the target environment, generate realistic social engineering content, identify high-value data, and evolve its tactics without human intervention. While fully autonomous AI malware has not been widely observed in the wild, AI tools are already being used to improve individual attack chain components.

History

Milestone	Year	Details
Concept	2000s	Self-adapting malware theory
DeepLocker	2018	IBM PoC; face-recognition trigger
AI phishing	2018	IBM: AI outperforms human phishing
WormGPT/FraudGPT	2023	Commercial LLMs, no guardrails
AI-enhanced phishing	2023+	Multiple APTs using LLMs

How It Works

Capability	Description
Targeted Phishing	LLMs generate personalised emails from OSINT
Vulnerability Discovery	ML scans code/config for unknown vulns
Evasion Adaptation	Observes detection → adjusts approach
Target Prioritisation	ML identifies valuable files/creds
Polymorphic Mutation	AI generates functionally equivalent variants
Social Engineering Bot	LLM conducts extended conversations

Notable Examples

Tool	Year	Notes
DeepLocker	2018	IBM PoC; hidden malware, face-recognition trigger
WormGPT	2023	Commercial; no safety guardrails
FraudGPT	2023	Commercial; fraud-focused LLM
AI phishing	2023+	APTs using LLMs for spear-phishing

Risk Assessment

Factor	Assessment
Current Maturity	Component-level (phishing, vuln discovery)
Fully Autonomous	Not yet observed in wild
Barrier to Entry	Lowering rapidly (open-source LLMs)
Defensive Gap	Signature-based AV ineffective

Defence

✅ AI-based defensive tools — Detect novel, adaptive patterns
✅ Robust email auth — DMARC, DKIM, SPF + AI phishing detection
✅ User training — Specifically on AI-enhanced social engineering
✅ Support AI safety research — Detection methods for adversarial AI
✅ Assume AI will be used — Defences not reliant on known patterns
✅ Behavioural detection — Not signature-based

MITRE ATT&CK

Technique	ID	AI Enhancement
Phishing	T1566	LLM-generated personalised content
Vulnerability Scanning	T1590	ML-assisted discovery
Obfuscation	T1027	AI-generated polymorphic variants
Social Engineering	T1598	LLM-driven conversation

Phishing — Social engineering
Polymorphic Malware — Evasion
Post-Quantum — Future threats