Attackers Are Using AI but Your Defense Is Still Waiting for Humans. That Asymmetry Ends Now

A finance employee at an engineering firm joined a video call to authorize a large wire transfer. Everyone on the call looked and sounded right, including the person presented as the company’s CFO. The transfer was approved, and the company was out $25.6 million.

The thing is: every participant on that call was a deepfake. Incidents like this are increasing quickly, and attackers are using AI to run operations that feel legitimate in real time.

Under these circumstances, companies have to reconsider how they approach cybersecurity.

The asymmetry that defines modern security

AI gives criminals unprecedented speed to hit their targets. Their tools run continuously, learning from every attempt and refining their methods without pause.

The economics of cyberattacks has shifted as well. AI removes the repetition that once made campaigns detectable, increases personalization, and produces content that blends into normal communication patterns.

Businesses need to respond to that shift, yet many still rely heavily on human-paced defense. Manual reviews, scheduled tests, and response processes activate only when an alert fires, and this imbalance leads to serious consequences.

The shape of an AI-powered attack

The range of techniques is wide nowadays. Criminals target identity, communication, infrastructure, and even the enterprise AI systems organizations deploy internally. The most active areas fall into four clusters.

AI-generated social engineering

You’ve received those clumsy, awkward messages with plenty of errors and poor formatting, haven’t you? Unfortunately, phishing attempts are now much harder to spot.

Criminals now use language models to draft emails that read as though a real person wrote them specifically for the recipient. AI helps them mine LinkedIn, public filings, and leaked databases for personal details that make messages credible. Each email has its own structure and thus looks different, which reduces the repeated indicators traditional gateways rely on.

Voice and video cloning raises the stakes further. A few minutes of audio from a public talk or an earnings call is enough to reconstruct someone's voice convincingly. Most people, including those who should know better, have no reliable way to tell a synthetic version from the real thing in the moment.

Also, business email compromise has moved past the single deceptive message entirely. Once attackers get into an account, AI can sustain an entire back-and-forth conversation with an employee, tuned to how that specific company routes approvals. This makes the resulting request far harder to question than a one-off email ever was.

AI-driven malware and exploitation

Malware has become far more adaptable. Some strains rewrite their own code at runtime, so each execution looks different to security tools. Others use external AI services to adjust their behavior on the spot, which makes static signatures much less useful.

Intrusions are also more automated. Certain tools can scan an environment, test defenses, and move through systems with minimal human involvement. Reconnaissance, exploitation, and lateral movement happen as one continuous flow.

Ransomware follows the same pattern. AI speeds up everything from initial access to full deployment, automating tasks like privilege escalation and payload preparation. The window for defenders to react becomes very small.

Identity and access targeting

Infostealers have become far more effective. Modern tools can collect credentials, session tokens, and browser cookies at massive scale. Machine-learning models then sort through these data dumps to identify high-value accounts tied to financial systems, cloud environments, or executive email.

Collaboration platforms are another target. Attackers impersonate IT help desks using legitimate Microsoft 365 tenants and run voice-phishing attempts through Teams. Since these messages come from trusted services and carry no obvious payloads, they’re hard to distinguish from normal internal communication.

AI-driven reconnaissance also plays a major role. Tools automatically gather social media activity, org charts, breach data, and public filings. This information is used in highly personalized spear-phishing and social-engineering campaigns that once required significant resources.

Attacks against enterprise AI

Malicious instructions can be embedded in an email, a document, or a calendar entry — anything an enterprise copilot is set up to read and act on. One compromised file can push an AI system to leak data or execute attacker-controlled actions. Traditional DLP tools often miss harmful content because it sits inside normal workflows.

Criminal groups also rely on unrestricted LLMs. Models like WormGPT 4 and KawaiiGPT generate phishing messages, scripts for moving through networks, and ransom notes with no guardrails. This lowers the skill barrier and allows less experienced actors to run campaigns that once required specialized expertise.

Why defense didn’t scale the same way

Attackers have automated both creation and execution, but defense largely stays manual. Alerts still wait for someone to review them, and suspicious requests still need a person to verify them. Judging whether something is a real threat or a false positive remains a human decision. These steps don’t compress the way automated reconnaissance does, because they were built around human attention as the core unit of work.

Governance gaps widen the problem. Many organizations are deploying internal AI tools — copilots, assistants, agents — without consistent access controls or oversight. Each ungoverned system becomes another surface an attacker can probe, at the exact moment their tooling is accelerating. The problem isn’t only that defense is slower. The area defenders must cover keeps expanding, often without anyone intentionally choosing it.

Closing the gap with AI-driven defense

Modern attacks move continuously, and to keep up, companies need an approach that works the same way. AI-driven continuous defense gives teams a real-time view of what's happening in their environment and helps them act before small issues turn into incidents.

An international energy company found out what this looks like the hard way. A government-backed group probed it for months, patient enough to wait for something to open. The SOC ran on signature-based detection, tuned to the specific traces known attack tools leave behind, and the group kept changing tools, so the traces kept changing with them. The alerts that should have fired never did.

What turned it around was a different basis for detection, and it maps to what continuous defense looks like in practice:

• Understanding real activity. Instead of matching fixed traces, rules can describe what a technique does to a system — where it reaches, how it moves — mapped to a known catalog of attacker methods like MITRE ATT&CK. Built this way, detection doesn't go stale when an attacker swaps tools, because it was never tied to the tool in the first place.
• Security checks that run with the workflow. Every change, deployment, or configuration update triggers a quick reassessment, so issues are found while the context is still fresh and easy to fix.
• Visibility into forgotten corners. AI maps areas that rarely get manual attention (old services, internal interfaces, or permissions never cleaned up). These quiet spots often become entry points.
• Response that doesn’t start from zero. Teams work from predefined playbooks. When an alert appears, they follow a familiar sequence instead of improvising under pressure. This shortens reaction time and keeps the process consistent.
• Governance for AI itself. As organizations adopt copilots and agents, they need clear rules for what these systems can do, monitoring for manipulation attempts, and response plans for model misuse.

Rebuilt this way, the energy company's SOC caught the same group's further attempts and contained them before persistence could take hold.

Conclusion

AI has significantly changed how attacks unfold. The only practical response is a defense model that works continuously, sees the environment as it really behaves, and reacts before attackers gain momentum.

Companies that adopt intelligent tools for constant monitoring and defense are able to operate at the same tempo as the threats they face.