The most useful way to think about AI in cyberattacks right now is as a throttle, not a steering wheel. Attackers are moving faster, hitting more targets, and spending less effort per intrusion. But their goals and their methods haven't changed much. Phishing, malware, and vulnerability exploitation still carry most of the load. What's shifted is how quickly those techniques can be chained together and how many times they can be tried.
That distinction matters because it determines where defenders should focus. If attackers had invented new techniques, the answer would be new detection categories. Instead, the answer is probably operational: shorten the time between a flaw becoming known and a flaw being closed, and reduce the number of places where a single weak credential or exposed service can open a path in.
The old playbook, run at machine speed
Ronald Lewis, director of cybersecurity governance at application security firm Black Duck, describes the attack lifecycle in familiar terms: attackers identify targets, perform reconnaissance, establish a foothold, move laterally, achieve their objective, cover their tracks, and move on. Nothing in that sequence is new. "What's changed is the volume and velocity that AI allows this process to happen," he says.
The most high-profile AI-enabled campaign so far fits that pattern. In September 2025, Chinese state-sponsored actors abused Claude Code to run an AI-orchestrated cyber-espionage campaign tracked as GTG-1002. The conventional techniques did the heavy lifting. AI served as an enabler for reconnaissance, vulnerability discovery, exploitation, lateral movement, credential harvesting, data analysis, and exfiltration. The tradecraft was familiar; the labor required to execute it was not.
JadePuffer, a ransomware campaign disclosed in July 2026, shows the same dynamic in a more destructive form. An LLM-driven agent automated most of the intrusion chain. The entry point was CVE-2025-3248, a missing-authentication flaw in Langflow, a popular open-source framework for building LLM applications and agent workflows. The flaw was already a year old when it was used. Internet-facing Langflow instances were looted for cloud credentials, API keys, and other secrets. From there, the attackers moved laterally, escalated privileges, and ultimately encrypted databases.
JadePuffer is useful as a case study precisely because it isn't exotic. The vulnerability was known, the exposure was visible, and the payoff came from automating a sequence defenders have seen for years.
Where AI gives attackers the biggest edge
If you're looking for the sharpest edge AI has given attackers, it's in finding flaws in software. AI tools can locate exploitable weaknesses in open-source code faster than most defenders can patch them. That's a serious problem for enterprises carrying years of unaddressed technical debt, where legacy components sit in production with known issues nobody has had the budget to fix.
Speed doesn't only help the winners. Darren Meyer, security research advocate at application security tools vendor Checkmarx, points out that "adversaries are realizing the same gains in development velocity and operations that everyone else is getting from implementing AI." Attackers can fail faster, too, which means they can discard dead ends and retry without the cost a human team would bear. Defenders are seeing gains in speed, scale, and operational ability that track closely with improvements in the underlying AI tools.
Lewis sees this expanding the attacker's reach. "An attacker can potentially investigate more opportunities, try more approaches, and move to the next target at a pace that would be difficult for a human team to match," he says. The practical effect is that opportunistic attacks become more thorough, and targeted attacks become cheaper to scale.
The dwell-time problem
This is where the model gets uncomfortable for defenders. Crystal Morin, senior cybersecurity strategist at cloud security firm Sysdig, notes that conventional intrusions have historically favored defenders because of dwell time. "That dwell time, which was often weeks or months, was where detection happened," she says.
Attackers used to land, sit quietly for a while, and poke around. That quiet period gave security teams a window to spot anomalies. AI-assisted attackers are showing signs of compressing that window, moving from foothold to objective far faster. And agents don't wait for a convenient moment to act. If your detection strategy depends on catching an intruder during a long, slow reconnaissance phase, that assumption is weakening.
Agentic APTs and the specialist problem
Human attack teams have always been specialized. One person handles web application exploitation, another writes malware, another understands cloud infrastructure. Agentic AI erodes that division of labor. An attacker using agents can draw on a broader base of domain knowledge, including database administration, CI/CD pipelines, and cloud configuration, without assembling a team to cover each one.
Nazar Tymoshyk, CEO and founder of agentic AI security vendor UnderDefense, puts it bluntly. "Think of it as a database admin, a developer, an IT engineer, a DevOps specialist, a cloud architect, and a hacker combined into a single role," he says. He calls the pattern "agentic APT," and argues the key change is that it's no longer limited to nation-state actors. "A threat that previously required a coordinated team of specialists can be deployed by a single person with the right prompt and the right model."
Tymoshyk also describes attackers switching models based on the job at hand. "One for reconnaissance, one for code generation, one for crafting social engineering content," he says. "The modularity is part of the attack design." UnderDefense and other researchers report that attackers favor Chinese models, specifically GLM and Kimi 3, which have significantly fewer content restrictions than their Western counterparts. Local models are also attractive for operational security reasons and because some may have weaker safeguards against misuse.
Meyer urges caution on that point. "Attribution of particular attacks to particular models is generally pure speculation from very thin evidence on the ground," he says. "True attribution can only come from infiltration of threat groups or compromise of threat group infrastructure, both of which are rarely available to public researchers." In other words, the model question is interesting but mostly unanswerable from the outside. Defenders can't build their response around which model a given attacker prefers.
The broader trajectory is clearer. Researchers describe a shift from AI-assisted and AI-accelerated attacks toward AI-orchestrated operations, where models surface, validate, and execute attack paths, and where human involvement shrinks to a minimum.
The economics are the real story
The cost structure of attacks is changing, and that may matter more than any single technical capability. Morin lays out the difference. "A human attacker builds a playbook per campaign and reuses it, so every new target costs engineering time to make adjustments," she says. "An agentic threat actor carries context about a whole class of applications and composes its plan live against whatever it finds. There, the only cost is inference budget. And as running a complex attack gets cheaper, the number of attempts will go up."
That's the uncomfortable math. When the marginal cost of trying another target approaches the price of compute, the economics favor volume. Defenders who rely on attackers being too expensive to bother with will find that assumption eroding.
That said, agentic attacks aren't flawless. Breadth of expertise is no substitute for depth, and agents make mistakes that give them away. Morin points to JadePuffer again: "Agents hallucinate and are noisy, which is often why they're catchable. Jadepuffer had flaws in its ransom note, encryption, wallet, and data exfiltration. There are still many areas in which human analysis and skill still outperform agent memory." Right now, the sloppiness of automated attacks is itself a detection signal, though there's no guarantee that gap persists as the tools improve.
What defenders have to change
Nathan Smolenski, CISO at vulnerability management and compliance firm Qualys, frames the problem around timing. "The core objectives of the threat actor haven't really changed, but the timeframe for a successful attack has been compressed dramatically," he says. "Frontier AI tools accelerate vulnerability discovery and weaponization, shrinking the window between disclosure and exploitation to near zero." His conclusion is direct: "Defenders can no longer rely on traditional, human-speed 30-day patch cadences or monthly static spreadsheets."
The entry surface is also widening. AI adds exposure through end-user devices, API calls, and ungoverned AI agents running inside organizations, often without anyone tracking what credentials they hold or what they can reach. The JadePuffer intrusion started with an exposed AI workflow tool, which is a warning worth taking literally.
The response, in Smolenski's view, is to move away from human-centric gatekeeping toward automated defenses, without abandoning the basics. Patching and authentication controls still matter, and arguably matter more when exploitation windows shrink. "Implementing agentic workflows with human-in-the-loop guardrails enables security teams to automate lower-level triage efforts and focus human expertise on governing agents, validating outputs, and making more critical risk decisions," he says.
My read is that the teams that come out ahead won't be the ones with the flashiest AI tools. They'll be the ones that inventory their AI and API exposure, shorten their patch cycles to match attacker speed, and let automation handle triage so analysts can focus on the judgment calls that agents still get wrong. The attacks haven't become unrecognizable. They've just become much harder to outpace by hand.