Cybersecurity in the Age of AI: New Threats and How to Fight Back

AI is upgrading cybercrime into faster, more targeted attacks that pressure your people and processes, and it is also introducing new technical failure modes inside LLM-based tools you deploy. You fight back by tightening identity and verification controls, hardening AI-enabled workflows, and running security operations with automation, logging, and response speed built in.

You will get a practical view of what is changing in phishing, fraud, malware, and nation-state activity, plus how to defend your organization without betting on “spot the deepfake” training. You will also get a clear operating model for using LLMs safely at work, with controls that map to real incidents and to widely used security guidance.

How Is AI Changing Cyberattacks Right Now (Phishing, Scams, Malware, And Nation-State Ops)?

AI is compressing the attacker’s cycle time. Recon, copywriting, translation, personalization, and multi-channel follow-up now run at scale with minimal human effort, so you face higher volume and higher relevance at the same time. That combination matters because humans fail when the message feels specific, urgent, and consistent with real business workflows. 

Defenders are also seeing more efficient attack chains that stitch together credential theft, infostealers, session hijacking, and rapid lateral movement. Microsoft reports processing massive security telemetry and highlights trends like AI-automated phishing and multi-stage chains that move faster than traditional manual operations. Your operational takeaway is simple: if detection and response still rely on slow handoffs, the attacker will keep winning time. 

AI also changes the economics of influence and deception. Deepfake audio and synthetic personas enable authority-based pressure without the usual “broken English” cues defenders relied on for years. You should treat that shift as a control problem: reduce what any single interaction can authorize, and force sensitive actions through authenticated, logged channels. 

What Are The Biggest AI-Powered Threats To Watch In 2025–2026?

Deepfake impersonation is now a mainstream business risk, not a novelty. It targets high-trust moments: help desk resets, executive approvals, vendor payment changes, and urgent operational exceptions. The threat is less about perfect audio fidelity and more about timing, context, and a request that exploits a gap in your verification process. 

AI-scaled phishing and messaging scams are the second major threat. Attackers now run coordinated outreach across email, SMS, voice, and encrypted apps, with consistent language and rapid iteration when you resist. Trend Micro has openly forecast scams that become AI-driven and AI-scaled, with emotional pressure tactics built into the playbooks, which aligns with what many security teams already see in payment fraud investigations.

The third category is “attacks on AI systems,” which many organizations underestimate until an internal chatbot becomes a data exposure channel or an automation trigger. OWASP’s Top 10 for LLM Applications documents repeatable issues that show up when LLM outputs are trusted too much, when plugins connect to real systems, and when prompt injection manipulates tool behavior. If you deploy internal copilots, customer-facing assistants, or AI agents connected to SaaS admin APIs, these risks become part of your core application security program.

Are Deepfake Voice And Video Scams Actually Happening, Or Is It Hype?

They are happening and they are operationally credible. The FBI issued a public alert dated May 15, 2025 describing an ongoing malicious text-and-voice messaging campaign where actors impersonate senior U.S. officials using AI-generated voice and smishing/vishing tactics to build rapport and then drive targets toward account compromise. That pattern mirrors what hits enterprises: a convincing identity claim, a quick relationship hook, then a pivot toward a link, a reset, a transfer, or a new messaging channel.

You should also notice the direction of targeting. These campaigns often go after people who can move money, access systems, or authorize exceptions, plus their assistants and help desks. That matches the reality that attackers do not need to break encryption if they can break workflow controls and identity checks.

Video deepfakes exist, yet voice is the more common operational weapon because it is cheap, fast, and fits into normal business behavior. A short voice note, a quick Teams call, or a voicemail can push urgency without triggering the scrutiny that a formal email request might. Your defense stance should assume that audio alone is not proof of identity.

How Do You Protect Your Business From AI Phishing, Vishing, And Deepfake Impersonation?

You protect the business by designing processes that stay safe even when a message is convincing. Training helps, yet it cannot carry the load when attackers can mimic tone, cadence, and internal vocabulary. Controls must force sensitive actions through strong authentication, strong authorization, and verifiable out-of-band confirmation.

Start by hardening the help desk and identity lifecycle. Token resets, MFA changes, SIM changes, password resets, and new device enrollments need step-up verification that an attacker cannot satisfy over a call. If the help desk can reset MFA from a phone request, you have built the attacker a bypass lane; tighten that policy, require authenticated portal workflows, and log every privileged identity change with clear ownership.

Then lock down financial and vendor change workflows. Vendor bank detail updates and payment approvals should require a two-channel verification rule using contact points sourced from your vendor master file, not from the request message. Add approval separation, minimum wait times for high-risk changes, and anomaly alerts when bank accounts or payment destinations change. These process-level controls remain effective even when the attacker uses perfect audio or flawless writing.

Last, remove speed as the attacker’s advantage. Verizon’s DBIR reporting has long emphasized that human-driven errors and social engineering play a major role in incidents, and the practical issue is fast compliance under pressure. Put friction into the few places where friction is healthy: resets, transfers, permission grants, external sharing, and emergency exceptions.

What Are The Top Security Risks Of Using LLMs At Work (Prompt Injection, Data Leaks, Agents, Plugins)?

LLMs introduce risks that blend classic application security with new failure modes in instruction-following systems. The most common business issue is sensitive data exposure: users paste customer data, credentials, internal incident notes, or proprietary code into tools without clear rules or technical guardrails. If that data flows into logs, analytics, browser history, or third-party processing, containment becomes difficult.

The next major risk is prompt injection and insecure output handling in systems that connect LLM output to actions. OWASP lists prompt injection as LLM01 and insecure output handling as LLM02 for a reason: if your tool copies LLM output into scripts, tickets, emails, firewall rules, or database queries without validation, the LLM becomes a conduit for malicious instructions. You should treat any LLM-generated output as untrusted input until it is validated, filtered, and constrained.

Agents and plugins raise the impact when you grant excessive permissions. OWASP flags insecure plugin design and excessive agency as recurring risks because an LLM that can click buttons in SaaS admin panels, call APIs, or execute workflows will eventually be manipulated into doing the wrong thing. The control here is not “tell the model to behave,” it is least privilege, scoped tokens, allowlisted actions, strong audit logs, and a kill switch that disables automation when anomalies show up.

Operationally, CISA and partners published joint guidance on deploying externally developed AI systems securely, explicitly focused on confidentiality, integrity, availability, mitigations for known vulnerabilities, and detection and response for malicious activity against AI systems and related data. Use that guidance to set baseline requirements for logging, patching, access, and monitoring before you let AI systems touch sensitive workloads.

Can AI Help Defenders More Than Attackers, And What Actually Works?

AI can meaningfully improve defense when you use it to accelerate triage, correlation, and response, and when you measure outcomes. It does not replace disciplined identity controls or sound engineering practices, and it will not save a SOC that lacks coverage, clear ownership, and tested playbooks. The winning pattern is “automation with guardrails,” tied to metrics that executives understand.

Cost and recovery data supports this when applied correctly. IBM’s Cost of a Data Breach Report 2024 reported a global average breach cost of $4.88M and noted that organizations using security AI and automation extensively saw $1.88M lower breach costs. The operational point is that detection speed and containment speed drive real money, and automation improves speed when it is integrated into incident workflows instead of sitting as an isolated dashboard.

Put AI where it can reduce toil and shorten time-to-decision. Use it for alert deduplication, enrichment, entity resolution, and drafting incident comms, while keeping approvals and irreversible actions under human control. Combine this with strong identity telemetry, endpoint visibility, and disciplined logging across SaaS and cloud control planes, since identity-based intrusion remains one of the highest-probability paths in modern environments.

What Guidance Should You Follow To Secure AI Systems In Real Organizations?

You need guidance that covers governance, engineering, and operations, and you need it mapped to your actual systems. NIST AI RMF 1.0 gives a structured way to govern and manage AI risk across the lifecycle with functions that help you organize ownership and measurement. Use it to set roles, risk appetite, documentation expectations, and testing requirements that survive executive turnover and vendor churn.

At the engineering level, OWASP’s LLM Top 10 provides a threat list that aligns with how LLM apps fail in production: injection, insecure output handling, supply chain issues, sensitive data disclosure, and excessive agency. Treat it like you treat OWASP Top 10 for web apps: design reviews, secure coding standards, pre-release testing, and targeted monitoring in production.

At the operations level, use CISA’s joint guidance on deploying AI systems securely to drive requirements for externally developed AI: supplier due diligence, configuration baselines, vulnerability management, and incident response coverage that includes AI-specific telemetry. If your vendor cannot support logging, access controls, and evidence during an incident, you are buying unmanaged risk.

How Is AI Changing Cybersecurity Threats—and How Do You Defend?

• New threats: AI-scaled phishing, deepfake voice scams, faster multi-stage attacks
• Core risk: Identity abuse, help desk resets, payment and vendor change fraud
• Best defense: Step-up verification, least privilege, logging, and fast automated response

Turn These Controls Into An Operating Rhythm That Holds Up Under Pressure

You do not win the AI era by chasing every new scam tactic, you win by tightening the few workflows that attackers exploit repeatedly. Lock down identity changes, make payment and vendor updates verifiable, and treat AI-connected tools like production software with threat modeling, testing, least privilege, and auditability. Use automation to shrink time-to-contain, and measure what matters: MFA coverage, patch latency, privileged access, mean time to detect, and mean time to respond. When these metrics improve, AI-enabled attackers lose the time advantage that makes their campaigns profitable. Build these controls into policy, tooling, and manager accountability so the defenses stay consistent even when the attacker’s content looks perfect.