The new cyber battlefield: How AI is changing both sides of the attack

Nearly 18 months ago, the UK’s National Cyber Security Centre (NCSC) made a bold prediction. It claimed that AI would “almost certainly” make cyber intrusions more effective and efficient, “leading to an increase in frequency and intensity of cyber threats”. It was right in this and virtually every other forecast detailed in the report.

But what it didn’t explore are the ways in which network defenders can also use the technology. If attacks are now moving at machine speed, the only way to stop them is with defensive approaches that match the same velocity and scale. As a new AI arms race emerges, it’s important that security teams seize this opportunity now.

Lowering the barriers to entry

AI has become a powerful tool for adversaries because is “compresses time and expertise”, according to Adam Ely, General Manager, AI Security at Check Point.

“Tasks that used to take a skilled team days, like researching a target, writing a convincing lure or analyzing code for weaknesses, can now be done in minutes by one person with the right model,” he says.

In short, AI isn’t necessarily rewriting the rules of the game but amplifying existing threats. It’s lowering the barriers to entry for would-be hackers while threatening to overwhelm network defenders still fighting the last war.

AI’s impact part one: vulnerabilities and phishing

At a high level, the use cases flagged by the NCSC are still relevant. Perhaps the most widely publicized malicious use of AI is in vulnerability research and exploit development (VRED). It’s empowering threat groups to find zero days in less time, in greater number and for lower cost than before.

It also enables them to pounce on newly published CVEs before defenders have had a chance to patch. Microsoft says “offensive timelines are increasingly measured in hours” rather than days or weeks. We suspect that timescale will soon drop down to minutes and seconds.

LLMs also help novice threat actors build convincing phishing campaigns in local languages. In fact, models are increasingly being integrated into pre-packaged offerings like phishing-as-a-service platform EvilTokens.

The impact is already being felt. One vendor caught a phishing email every 19 seconds in 2025, more than double the rate a year previously. And it’s not just volume increasing: Microsoft claims AI-automated phishing emails achieved a 54% click-through rate, versus 12% for standard attempts.

AI’s impact part two: reconnaissance and malware generation

AI is also helping to streamline reconnaissance; rapidly sifting through publicly available information on potential targets, mapping IT environments and scanning for exposed endpoints. By doing the heavy lifting, it frees up threat actors to focus on higher-value work.

AI is getting better at basic malware generation too – what is effectively malevolent vibe coding. In more advanced examples, like VoidLink, the AI actually communicates with the malware while it’s running on a victim’s machine to generate new commands or rewrite code on the fly.

Rise of agentic AI

Agentic AI gives threat actors yet another opportunity: this time to orchestrate attacks. Anthropic disclosed one such campaign in June, claiming Claude Code handled 80-90% of reconnaissance, exploitation, credential harvesting, lateral movement, and data triage across 30 target organizations.

Google more recently revealed how a hacker used an AI coding chatbot and a set of agent instructions to plan, build and execute a mass credential harvesting campaign in under six hours. It claimed the agent autonomously managed the vulnerability scanning pipeline, performed real-time troubleshooting, and executed IP rotation logic without manual intervention.

From frontier models to open weights

Much of the innovation in this space comes in the AI supply chain, rather than the underlying attacks, explains Adrian Cheek, Senior Cybercrime Researcher at Flare. He cites infostealers designed to capture AI provider keys and session tokens from developer machines. These can subsequently be used to run attack workloads at someone else’s expense, he says.

“AI agents on a developer’s laptop may be configured to reach code repositories, cloud consoles, databases and ticketing systems,” Cheek explains. “When a stealer compromises that machine, the agent’s configuration can reveal where valuable secrets live, and the secrets themselves may already be accessible from the same environment. In that situation the AI key points to everything else.”

An alternative to hijacking paid-for AI models is to turn to publicly available versions.

“Both open-weight models and illicit access to commercial models will narrow the gap between frontier capability and what criminals can use,” he warns. “Defenders should assume that capabilities demonstrated by the most capable models will eventually reach less sophisticated actors.”

A prevention-first approach

Fortunately, security teams also have options, although not all are as effective as each other. AI can help network defenders analyze billions of signals at machine speed to improve detection and response. But the impact can be limited, says Check Point’s Ely.

“Attackers only need to succeed once, while defenders have to be right every time. That asymmetry is why defenders have to use AI to prevent attacks, not just detect them after the fact,” he argues. “When AI can move from vulnerability to breach in minutes, detect and respond alone is too slow.”

This should put the focus on prevention-first efforts which stop attacks before they cause any damage, such as using AI tools to assist with vulnerability and patch management. Ely advises organizations to continually manage exposure to “fix what attackers will find first”, and to include AI infrastructure in any prevention-first strategy, as it’s now “part of your attack surface”. A unified platform is also important, to improve visibility and help enforce a single policy across diverse systems.

“Check Point’s ThreatCloud AI engines analyze threat intelligence from across our global network to block threats in real-time, across network, cloud, workspace and AI,” Ely explains. “We launched our Frontier AI Readiness Program to prepare customers for the next generation of AI-discovered vulnerabilities. And we work with frontier AI labs to bring their most advanced cyber models into our platform.”

The arms race goes agentic

There’s plenty more road left to run. Google predicts the emergence of fully autonomous “exploitation pipelines” with agents orchestrating every stage of an attack, for example. Keyfactor CSO, Chris Hickman, agrees that agentic AI will be where threat actors gravitate towards.

“As models learn more and AI agents begin to leverage that knowledge, threats will become more sophisticated and therefore harder to detect,” he says. “One of the greatest challenges for defenders will be to decipher legitimate AI identity and behavior from nefarious agents.”

To do so, they must adapt approaches like zero trust to rethink assumptions around static identity and trust, and to ensure security can operate at the same scale, speed and scope as AI. “We are still in the infant years of the AI arms race,” Hickman adds.

For security leaders overwhelmed by the prospect of all the battles yet to fight, there are some words of cautious encouragement from Check Point’s Ely.

“The good news is that defenders get the same technology, and we have advantages attackers don’t: visibility into our own environments, massive, shared threat intelligence, and the ability to put controls in place before an attack,” he concludes. “The organizations that fall behind will be those still relying on manual processes and point products.”