Seven Principles Behind the Adoption of AI in Cybercrime
Key Takeaways
- AI 在網路犯罪中的應用具有選擇性,並非所有攻擊都能自動化 網路犯罪者主要將 AI 應用於網路釣魚、詐騙支援、多語言翻譯、程式碼撰寫輔助、目標偵察及被竊資料分類等大量、重複且失誤成本較低的工作。相較之下,橫向移動、權限提升及維持對受害系統的存取權等需要長時間掌握攻擊脈絡與判斷能力的複雜入侵活動,目前仍較難可靠地自動化。
- AI 降低網路犯罪的人力需求,讓小型犯罪團隊也能擴大攻擊規模 AI 並未取代人類犯罪操作者,而是減少原本需要多人分工的詐騙對話、多語言翻譯及程式碼撰寫工作。小型犯罪團隊因此能同時產生更多內容、接觸不同語言的受害者,並加快被竊資料的分析與變現,降低經驗較少的犯罪者執行犯罪活動的門檻。然而,複雜的入侵行動仍仰賴具備專業經驗的攻擊者。
- AI 代理程式已參與實際入侵,但自主攻擊能力仍受到限制 研究觀察到 AI 代理程式(AI Agent)已在實際攻擊活動中協助目標偵察、取得憑證、建立通道及資料外傳等操作,但人類操作者仍掌握控制權,且相關活動受到技術可靠性、基礎設施存取權及人為監督等條件限制。 另一方面,AI 驅動惡意程式仍屬漸進式發展,尚未形成可自給自足的生態系。企業應優先強化身分與存取控制、行為式威脅偵測、資料暫存階段偵測及自動化防禦。
From Capability to Adoption
Discussion of AI-enabled cybercrime often starts with capability (deepfakes, automated phishing, malware generation, autonomous hacking), but criminals don’t adopt technology because it’s impressive in isolation. They adopt it when it improves return, reduces friction, fits existing workflows, and carries acceptable operational risk. The evidence available in 2026 points to selective adoption rather than uniform transformation:
| Uses of AI in cybercrime that have stabilized | Emerging uses of AI in cybercrime |
|---|---|
| Phishing content, fraud support, localization, coding assistance, reconnaissance, stolen-data processing | Agentic intrusion*, AI-powered malware* (*constrained by reliability, access, infrastructure dependency, and the need for human control) |
In our report, “Seven Principles Behind the Adoption of AI in Cybercrime,” TrendAI™ Research organized criminal AI adoption around seven principles that explain why AI is already useful when it improves the return on an operation, reduces friction, fits existing workflows, but fully autonomous intrusion and self-sufficient AI malware are emerging but bounded. Read more of our research here.
Principle 1: Industrialization, not novelty or autonomy
Autonomous AI-powered hacking has not been observed in practice. Criminals instead use AI to speed up existing workflows: lures, scam translation, stolen-data parsing, code, fraud conversations. The effect is industrialization: more volume and variation, faster adjustment to signals, concentrated wherever it reduces friction.
Principle 2: AI is strongest where failure is cheap
AI gains traction fastest where output can be generated, corrected, or discarded cheaply at volume: phishing copy, fake support chats, translation, credential-log parsing, stolen-data classification. Imperfect output is still useful here, and phishing content has gotten sharper in grammar, tone, and localization, weakening the tells that once helped users spot scams.
Principle 3: AI is weaker where continuity matters
Long intrusion chains (lateral movement, persistence, privilege escalation, orchestration) need context and judgment AI handles less reliably. A mistake inside a defended network can burn access, unlike a correctable mistake in a phishing draft. Even as frontier models (e.g., Anthropic’s Mythos-class systems) show stronger vulnerability-discovery skill, that access stays restricted and monitored, so criminal adoption depends on cost and durability, not raw capability alone.
Principle 4: Smaller criminal teams, more productivity
AI compresses labor once spread across conversationalists, translators, and script writers, letting one operator, or an agent, run more conversations, languages, and code simultaneously. It doesn’t remove human operators; it raises the operational floor for lower-tier actors, while high-end intrusion stays dependent on experienced ones.
Principle 5: Faster monetization of stolen data
Attackers often steal more data than they can manually evaluate. AI classifies stolen files, surfaces high-value accounts, and shortens the path from access to monetization, accelerating the shift toward modular monetization beyond single-encryption ransomware.
Principle 6: Services, not superintelligence
The underground AI market isn’t led by dramatic “super-AI” tools. Brands like WormGPT, Xanthorox, FraudGPT, DarkGPT, and EvilGPT market unrestricted capability, but mostly rely on repackaged commercial-model access and jailbreak workflows. Integration into phishing kits and fraud platforms, not branding, is the real signal of staying power.
Principle 7: Agentic intrusion is real, but bounded
AI agents are now assisting reconnaissance, credential access, tunneling, and exfiltration in live operations. TrendAI™ Research’s SHADOW-AETHER-040 and SHADOW-AETHER-064 observations show agents assisting while actors stayed in control. JADEPUFFER went further, chaining exploitation of an exposed Langflow instance into destructive database extortion, but along a path enabled by weak configurations, not full autonomy. AI-powered malware (PromptLock, LAMEHUG, MalTerminal) shows similar incremental movement, not a self-sufficient ecosystem.
Read more on the criminal AI underground in our research.
Defensive priorities
| Identity and access controls become more important as AI compresses the time between compromise, triage, and monetization. | Behavior-based detection becomes more important as AI improves content quality; better phishing text doesn’t make delivery infrastructure legitimate. |
| Staging detection becomes more important as AI improves stolen-data triage: watch for unusual file enumeration, selective data movement, and bulk archiving. | Defensive automation becomes more important as attack tempo increases, reserving human judgment for the decisions that matter most. |
Criminal AI is reshaping cybercrime unevenly: strongest where failure is cheap and tasks are repetitive, weaker where operations require continuity and judgment. Agentic intrusion and AI-powered malware are real developments, but remain bounded by reliability, infrastructure dependency, and human supervision, for now. The near-term impact isn’t full autonomy, but industrialization: familiar criminal workflows becoming faster, cheaper, and easier for smaller teams to operate, shrinking the time between access and monetization. The future of criminal AI will be defined less by what the most powerful model can do, and more by what criminals can access, package, integrate, repeat, and monetize.
Explore more of our research into how AI is reshaping the cybercrime landscape by reading the full report.