Security teams spent 2024 debating whether AI would change threat landscapes. In 2026, the answer is clear: it already has. AI has not created fundamentally new attack categories — phishing is still phishing, malware is still malware — but it has dramatically lowered the skill floor and increased the speed, scale, and personalization of attacks. Here is what is actually happening and what defenses are working.
AI-Generated Phishing: Volume and Personalization at Scale
Traditional phishing campaigns were limited by the effort required to write convincing emails. LLMs removed that constraint. In 2025, IBM X-Force documented a 112% increase in AI-assisted phishing emails. These messages are grammatically perfect, contextually relevant, and can be personalized at scale using data scraped from LinkedIn, company websites, and breach databases. The result: click rates on AI-generated phishing are approximately 2x those of traditional campaigns, according to SlashNext research.
Specific patterns observed in the wild:
- AI-generated spear-phishing targeting finance teams with CFO impersonation and invoice fraud language
- Voice cloning attacks (vishing) using 3-second audio samples from public videos to impersonate executives
- AI-crafted pretexting scripts used in BEC attacks to manipulate wire transfer approvals
- Deepfake video calls used to impersonate IT support and request credential resets
Verified in 2025: A UK engineering firm lost $25 million after attackers used a deepfake video call to impersonate their CFO and request an emergency wire transfer. The impersonation was convincing enough that four different employees approved the transaction.
AI-Assisted Malware Development
LLMs lower the skill floor for malware development. Security researchers demonstrated in 2024 that GPT-4 could generate functional polymorphic malware — code that rewrites itself to evade signature detection — when given sufficiently indirect prompts. By 2026, darknet forums sell purpose-built “uncensored” LLMs trained specifically for offensive security use. These tools help less-skilled attackers write credential stealers, ransomware droppers, and C2 frameworks without deep programming knowledge.
Context: AI-assisted malware is not yet significantly more sophisticated than human-written malware — it is faster to produce and more accessible to lower-skill attackers. The bigger risk is volume: more threat actors can launch more campaigns simultaneously.
Automated Vulnerability Discovery
AI-powered fuzzing and vulnerability discovery tools are accelerating the time between CVE publication and exploit availability. Google DeepMind’s BigSleep project discovered a zero-day vulnerability in SQLite using AI-assisted analysis in October 2024. Nation-state actors are deploying similar capabilities at scale — meaning that when a critical vulnerability is published, exploitation attempts begin within hours, not weeks.
This compresses the acceptable patch window for internet-facing systems from days to hours. Your vulnerability management program needs to account for AI-accelerated exploitation timelines.
Prompt Injection: The New Attack Surface
Every organization deploying LLM-powered tools — AI assistants, chatbots, document summarizers, email helpers — has introduced prompt injection attack surface. Prompt injection occurs when malicious text in user input or processed documents manipulates the AI model’s behavior: leaking system prompts, exfiltrating conversation history, bypassing access controls, or causing the AI to take unintended actions.
In agentic AI systems (AI with tool access and the ability to take actions), prompt injection can result in real-world consequences: sending emails, accessing files, executing code, or making API calls the user never authorized. This is not a theoretical risk — security researchers demonstrated prompt injection attacks against GPT-4-based email assistants, Bing Chat, and enterprise AI copilots throughout 2024 and 2025.
AI-Powered Defenses That Are Actually Working
Behavioral Detection in EDR and SIEM
Modern EDR platforms (CrowdStrike Falcon, Microsoft Defender XDR, SentinelOne) use ML models trained on billions of endpoint events to detect anomalous behavior without signature matching. This matters because AI-generated malware bypasses signature detection but still exhibits detectable behavioral patterns: process injection, credential access, network beaconing, and lateral movement.
AI-Powered Phishing Detection
Email security platforms like Abnormal Security and Darktrace Email use behavioral AI to detect phishing that bypasses link and attachment scanning: account takeover via legitimate compromised accounts, BEC with no malicious links, and AI-generated content that evades keyword filters.
Copilot for Security (Microsoft) and Charlotte AI (CrowdStrike)
Both platforms use LLMs to assist SOC analysts: summarizing incidents, explaining attack chains, suggesting investigation steps, and drafting response actions. Early deployments show 26-40% reduction in mean time to respond for organizations with mature security programs. These tools do not replace analysts — they reduce the cognitive load on experienced analysts handling high alert volumes.
What Your Security Team Should Do Now
- Update phishing training to include AI-generated email examples and deepfake voice scenarios
- Implement DMARC, DKIM, and SPF on all sending domains — AI phishing still needs a convincing sender address
- Adopt a verification protocol for wire transfers and credential requests that does not rely on voice or video calls alone
- Audit all AI tools deployed in your organization for prompt injection attack surface and data leakage risk
- Compress your patch window for critical CVEs to 24-48 hours for internet-facing systems
- Evaluate behavioral AI detection tools in your EDR and email security stack — signature-only detection is insufficient in 2026
Takeaway: AI has changed threat timelines and lowered attacker skill requirements. The defenses that matter most in response — MFA, behavioral detection, rapid patching, and employee awareness training updated for AI threats — are not new. The urgency is. Subscribe to our weekly newsletter for ongoing AI threat intelligence coverage.
n
🔗 Further Reading
Frequently Asked Questions
n
How is AI being used in cyberattacks in 2026?
n
AI is used in four primary attack categories: (1) AI-generated phishing — LLMs create personalized, grammatically perfect emails at scale, achieving 2x higher click rates than traditional phishing. (2) Voice cloning — 3-second audio samples enable real-time executive impersonation in vishing attacks. (3) Deepfake video — used in BEC attacks to authorize fraudulent wire transfers. (4) AI-assisted malware — purpose-built offensive LLMs help low-skill attackers write credential stealers and ransomware droppers.
n
n
What is prompt injection in AI security?
n
Prompt injection is an attack where malicious text in user input or processed documents manipulates an AI model's behavior — causing it to leak system prompts, bypass access controls, exfiltrate data, or take unintended actions. In agentic AI systems with tool access (ability to send emails, access files, execute code), prompt injection can have real-world consequences. It is the most significant new attack surface introduced by enterprise LLM deployments in 2025-2026.
n
n
How can organizations defend against AI-powered phishing?
n
Defense against AI-generated phishing requires multiple layers: deploy DMARC, DKIM, and SPF on all sending domains; use AI-powered email security (Abnormal Security, Darktrace Email) that detects behavioral anomalies rather than relying on link/attachment scanning; update security awareness training with AI-generated phishing examples; implement a verification protocol for wire transfers and credential requests that requires a separate, out-of-band confirmation channel not vulnerable to impersonation.
n
n
What is the NIST AI Risk Management Framework?
n
The NIST AI Risk Management Framework (AI RMF 1.0) is a voluntary framework published in January 2023 to help organizations manage risks associated with AI systems. It is organized around four core functions: Govern (establish AI risk management culture and policies), Map (identify and categorize AI risks), Measure (analyze and assess AI risks), and Manage (prioritize and address AI risks). US federal agencies and many private sector organizations now require alignment with the AI RMF for AI deployments.
n
n
n
