The Industrialization of Cyberattacks

AI is doing to cyberattacks what the assembly line did to manufacturing

Throughout history, industrialization has turned activities that depend on individual craftsmanship into standardized and repeatable production processes that can operate at scale. Its economic significance comes from reducing the cost of producing each additional unit by replacing expensive human labor with machinery, automation and specialized infrastructure, yet the consequences of those lower costs reach far beyond productivity gains because they change who can participate in an industry, how production is organized and what becomes economically viable.

Cyberattacks have been moving in this direction for decades. Botnets automated attacks against enormous numbers of machines while exploit kits packaged vulnerabilities into reusable products and ransomware-as-a-service allowed criminals to specialize in different parts of an operation. Yet despite these developments sophisticated cyber operations remained constrained by one particularly expensive and scarce resource: human expertise.

Discovering vulnerabilities, developing exploits, navigating unfamiliar infrastructure and adapting to unexpected results require technical knowledge and cognitive abilities that are expensive to acquire and difficult to replicate. For decades, this scarcity placed a natural limit on the scale of sophisticated cyber operations.

Artificial intelligence is beginning to remove that constraint because increasingly capable models and the falling cost of intelligence allow work that once required skilled human operators to be delegated to software that runs continuously and reproduces it at scale.

The consequences extend beyond making individual hackers more productive. Lower costs change which targets are worth attacking, who can afford to conduct sophisticated operations, and how the cybercrime industry itself is organized.

We are witnessing the industrialization of cyberattacks.

From human expertise to compute

The industrialization of cyberattacks did not begin with artificial intelligence.

Tools like Nmap automated network discovery, Metasploit made exploitation more accessible, and Cobalt Strike simplified the coordination of complex offensive operations. Each generation progressively transferred another technical capability from the human operator into software, making it easier and cheaper to perform activities that previously required specialized knowledge.

Figure 1 From automating execution to automating intelligence Each generation of security tooling progressively transferred technical work from human operators into software. AI is beginning to automate the reasoning that connects those operations.

1990s

Network discovery, analysis and scanning

Discovering hosts, identifying services, inspecting traffic and checking for known vulnerabilities.
- SATAN , opens in a new tab 1995
- Nmap , opens in a new tab 1997
- Nessus , opens in a new tab 1998
- Wireshark , opens in a new tab 1998
2000s

Exploitation frameworks and web testing

Reusing known exploits, executing payloads and testing web applications for common vulnerabilities.
- Core Impact , opens in a new tab 2002
- CANVAS , opens in a new tab 2002
- Metasploit , opens in a new tab 2003
- Burp Suite , opens in a new tab 2003
- sqlmap , opens in a new tab 2006
2010s

Command and control and post-exploitation

Managing sessions, coordinating post-exploitation tasks and operating across compromised networks.
- Mimikatz , opens in a new tab 2011
- Cobalt Strike , opens in a new tab 2012
- PowerShell Empire , opens in a new tab 2015
- BloodHound , opens in a new tab 2016
- Sliver , opens in a new tab 2019
Late 2010s to 2020s

Automated adversary emulation

Running predefined attack sequences, reproducing known adversary techniques and validating controls.
2023–2025

AI-assisted research

Reasoning through reconnaissance, vulnerability discovery and multi-step tests, in research and early products.
- PentestGPT , opens in a new tab 2023
- PentAGI , opens in a new tab 2025
- Strix , opens in a new tab 2025
- ARTEMIS , opens in a new tab 2025
2026 onwards

Operational AI agents

Planning and running whole operations, in documented real-world attacks.
- HexStrike AI , opens in a new tab 2025
- CyberStrikeAI , opens in a new tab 2025
- ARTEX , opens in a new tab 2026
- CAIRN , opens in a new tab 2026
- Apex , opens in a new tab 2026

Source: release dates from each tool's repository, documentation or launch announcement, linked from its name.

However, these tools had an important limitation. Although they could execute increasingly sophisticated operations, humans were still responsible for understanding the environment, interpreting results, identifying opportunities and deciding what to do next.

The tools and operators were complementary inputs, where making one input cheaper increases productivity, but the scale of production remains constrained when the other input is scarce. A thousand automated scanners provide limited additional value if only a handful of skilled operators are available to investigate their findings and turn them into successful attacks.

AI is beginning to change this relationship because it can substitute for portions of that human cognitive work rather than simply complementing it.

Figure 2 Output now scales with compute, not with headcount Finding vulnerabilities now grows with the compute you spend, not with the experts you can hire.

Source: Cantina, What We Learned Hunting for Bugs in Over 1500 Codebases with AI (2026), opens in a new tab.

Modern models can analyze source code, formulate hypotheses, interact with tools, interpret results and adapt their approach across multiple steps. When combined with agentic frameworks that provide execution environments, memory and verification mechanisms, activities that previously required continuous human involvement can increasingly be delegated to software.

The economic implications become particularly interesting when considering how quickly the price of intelligence is falling. As models become more capable and inference costs decline, the same computational resources can perform increasingly sophisticated work that previously required expensive human expertise.

This changes the cost structure of offensive operations. While hiring and training skilled operators requires significant time and capital, software can be reproduced, executed continuously and scaled across thousands of tasks at a marginal cost increasingly determined by compute. The constraint begins to shift from how many skilled operators an organization can employ to how effectively it can deploy and coordinate autonomous systems.

Figure 3 The changing economics of cyberattacks As artificial intelligence becomes cheaper and more capable, offensive security is shifting from an industry constrained by human expertise towards one increasingly driven by scalable computation.

The expertise that once capped offense can now be bought as compute

Source: Cantina, What We Learned Hunting for Bugs in Over 1500 Codebases with AI (2026), opens in a new tab.

Within months, anyone can download frontier cyber capability and run it without safeguards

170

GPT-6 Astra No open model yet

160

4 mo Claude Mythos Preview GLM-5.3

150

6 mo Claude Haiku 4.5 GLM-5.1

140

4 mo o3 gpt-oss-120b

130

7 mo Claude 3.5 Sonnet DeepSeek-R1

202420252026

First closed and first open-weight model to reach each Cyber ECI level
Cyber ECI level	First closed model	First open-weight model	Lag
170	GPT-6 Astra, Sept 2026	None yet	Not yet matched
160	Claude Mythos Preview (April), Apr 2026	GLM-5.3, Aug 2026	4 months
150	Claude Haiku 4.5, Oct 2025	GLM-5.1, Apr 2026	6 months
140	o3, Apr 2025	gpt-oss-120b, Aug 2025	4 months
130	Claude 3.5 Sonnet, Jun 2024	DeepSeek-R1, Jan 2025	7 months

Sources: Epoch AI, Cyber ECI (November 2026), opens in a new tab; Artificial Analysis, Cyber Index refusal rates (November 2026), opens in a new tab.

Of course, the relevant economic variable is not how cheap a model is to run, but how much it costs to produce a useful result. An autonomous agent that spends hours investigating a vulnerability without finding anything may be less productive than an experienced researcher who identifies it in minutes.

But as model capabilities improve, the probability of completing increasingly complex tasks rises, while the amount of human involvement required to perform them can fall. The result is a reduction in the effective cost of producing offensive work, provided those productivity gains outweigh the cost of computation and unsuccessful attempts.

When our own exploit factory, opens in a new tab hunts for vulnerabilities, its output follows a predictable production function, opens in a new tab rather than the uneven results of individual talent. Doubling the compute adds about 40 percent more findings, the same curve holds from one codebase to the next and identical runs vary by only five to six percent. Measured against human penetration tests that cost more than $500,000 and ran for months, it recovered every critical and high-severity vulnerability the testers had found for a fraction of the compute the same budget would buy. Expertise that was scarce, slow and unpredictable is becoming a standardized input whose output can be planned in advance, which is precisely what industrialization does to work that once depended on individual skill.

And when the cost of production falls, the economic consequences are not limited to productivity.

Mass production

Every cyberattack involves an economic decision, whether that decision is made explicitly by an organized criminal group or implicitly by an individual choosing what to target.

An attacker must invest resources in discovering an opportunity, acquiring the necessary capabilities, conducting the operation and extracting some form of value. An operation becomes economically attractive when the probability-weighted value of success exceeds its expected cost.

This means that not every vulnerable system is necessarily worth attacking. A small organization may be technically easy to compromise, but if the expected return is low and the operation requires several days of an experienced operator's time, the opportunity may simply not justify the investment.

The economics change in two ways when the same work can be delegated to an autonomous system at a fraction of the cost:

Attackers can pursue existing opportunities more cheaply, increasing the potential return of successful operations.
Opportunities that were previously uneconomical become worth investigating because the marginal cost of exploring another target has fallen.

Figure 4 Lower attack costs expand the economically viable target pool As AI reduces the cost of offensive work, opportunities that previously weren't worth pursuing can become economically attractive.

E [Π_{i}] = p_{i} V_{i} - C_{i}

viable when

p_{i} V_{i} \geq C_{i}

p is the probability of success, V the value of a successful operation (financial for criminals, strategic for state-sponsored operations) and C the total expected cost of attempting it, including human labor, inference, infrastructure and operational expenses, not only model tokens.

An internal application, a forgotten staging server or a small supplier that was not worth an expert's attention may suddenly become worth a few minutes of automated investigation. Low-severity vulnerabilities are now worth exploiting too because a model can match a leaked software version against public disclosures and build a working exploit, which makes severity a question of how cheaply a flaw can be exploited.

This is one of the most important consequences of industrialization. When production costs fall dramatically, the effect is not limited to producing the same quantity with fewer resources. Lower costs expand the set of economically viable opportunities, attract new participants and enable levels of production that were previously impossible.

The same reasoning explains why automation can increase the total volume of activity even when the value of each individual operation is relatively small. An operator who previously needed to choose carefully between ten potential targets may now have the resources to investigate thousands, allowing unsuccessful attempts to be absorbed into the operating cost of a much larger process.

The result is a change in the economics of target selection. Instead of concentrating exclusively on opportunities that justify significant human effort, attackers can increasingly explore a much broader distribution of targets, prioritizing those that produce the highest expected return relative to their computational and operational costs.

Exploit factory

The falling cost of intelligence is not only changing which targets are economically attractive but also how offensive operations are organized. Throughout 2025 and 2026 AI evolved from assisting individual operators with coding, research and other discrete technical tasks towards agentic frameworks capable of planning, executing and adapting multiple stages of an attack, progressively reducing the human involvement required to coordinate increasingly complex operations, and by late 2026 open-source frameworks such as ARTEX, opens in a new tab, PentAGI, opens in a new tab and CAIRN, opens in a new tab were already being used in real-world attacks.

Figure 5 The evolution of autonomous offensive frameworks From AI-assisted tasks to multi-stage autonomous intrusions in less than a year.

Stage 1: Prompt-based use of LLMs

End of 2025 to early 2026

Adversaries use LLMs like a chatbot to write and debug malware, draft phishing emails, create personas and run light reconnaissance.
Stage 2: LLMs embedded in malware

Early 2026

Malware such as HONESTCUE and PROMPTSPY calls LLMs during execution, for example to obfuscate payloads and evade defenses.
Stage 3: LLM-enabled automation in operations

Early 2026 to September 2026

Agents handle discrete tasks and work alongside human operators on novel exploits and high-stakes vulnerability research.
Stage 4: Experiments with autonomous frameworks

Mid 2026 to September 2026

Campaigns offload more stages of the attack lifecycle to LLM-augmented tooling and wire model APIs into tools and bots.
Stage 5: LLM-driven intrusions in the wild

Current stage, October 2026

LLM-driven tooling conducts multiple stages of real intrusions with minimal human intervention, messy and noisy yet successful.

Source: Google Threat Intelligence Group and Mandiant, ARTEX and Friends Enter Agentic Offensive Frameworks (October 2026), opens in a new tab.

Seen this way an autonomous framework works like an assembly line. The attack lifecycle is broken into stations for reconnaissance, vulnerability discovery, exploitation and post-compromise work, each staffed by specialized agents that run in parallel and pass their results down the line while a human sets the production target and inspects what comes out. The inputs are compute, models, tooling, vulnerability intelligence and a list of targets, and every one of them is getting cheaper. The outputs are access, stolen credentials, exfiltrated data, working exploits and stolen funds, alongside a large volume of failed attempts that the line absorbs as scrap.

Figure 6 Anatomy of an autonomous attack factory An emerging architecture for industrial-scale offensive research, combining parallel autonomous workers, persistent knowledge, shared execution infrastructure and continuous verification.

Control plane

Human supervisor objectives, authorization, budgets, approval gates
Planner / orchestrator decomposes tasks, manages dependencies, replans from results
Durable task queue pending and active jobs, retries
Persistent scheduler recurring investigations, task lifecycle

↓ Workers claim tasks from the queue asynchronously

Execution plane, parallel

Discovery workers × N
Research workers × N
Validation workers × N
Analysis workers × N

Shared by every worker: model inference (frontier and open-weight models), tool execution behind an approval gate, and isolated environments holding authorized test targets only.

↕ Workers and planner read and write shared state

Shared data plane, persistent

Asset graph the environment: assets, services, dependencies
Exploration graph what was investigated: hypotheses, paths, findings, linked to assets by anchors
Memory and state store working context, task state, history
Artifact store logs, outputs, findings, evidence

↓ Candidate results go to verification

Evaluation plane, continuous

Verification engine reproducibility checks, deterministic tests, evidence validation
Evaluation and scoring validity, confidence, duplicates, outcome classification

↺ Feedback to the planner: verified results, failed paths, new research tasks

An attack factory runs on the economics of a production line, where the cost of each output is the cost of the inputs divided by the yield, meaning the share of attempts that actually produce something usable. Compute and models keep getting cheaper, parallel agents raise throughput without adding operators, and the infrastructure, tools and knowledge built for one campaign carry over to the next so that fixed development costs are spread across far more attempts. Yield remains the weak point because most attempts fail somewhere along the line, whether the target turns out not to be vulnerable, the exploit breaks or a defense stops the intrusion, yet a low yield matters less when each attempt costs so little, failures are retried automatically and the line can simply run more of them. Defense works on the same equation, since MFA, fast patching and endpoint detection lower the yield and raise the cost of each attempt, while takedowns of criminal infrastructure force fixed costs to be paid again.

Figure 7 The unit economics of an attack factory Shared infrastructure and cheaper computation let offensive work scale, lowering the cost of each success as long as yield holds up.

Risk

Defenses p↓ c↑ MFA, patching and EDR lower the yield p and raise c
Takedowns F↑ Seized servers and blocklisted domains force F to be paid again

Costs

Fixed, paid once per campaign F

Tooling Frameworks adapted once
Setup Servers and orchestration stood up once
Supervision One operator, whatever the volume
Knowledge Carried from one campaign to the next

Variable, paid on every attempt c

Inference Tokens from closed or open models, cheaper every year
Consumables Proxies, accounts and domains burned

Raw materials close to free

Intelligence Public CVE and KEV data
Targets Lists of exposed services

The line

N attempts enter; the funnel shows how many are still in play at each stage

Recon
Discovery
Exploitation
Post-compromise

Yield p: the share that reaches a defined objective; the rest are scrap

Failures feed back: agents read the error and retry

Outputs, each worth V

Access Sold to brokers or used for ransomware
Credentials Resold or reused
Data Extortion or sale
Exploits Reused on the next target
Funds Ransoms and stolen crypto

Reinvested in the line

Cost per success $= \frac{F + N c}{N p}$

An attempt is worth making when $p V \geq c + \frac{F}{N}$

F Fixed capital: tooling, setup and knowledge, paid once per campaign
N Production volume: attempts, which grow with parallel agents rather than operators
c Marginal cost: inference and consumables per attempt, falling every year
p Yield: the share of attempts that reach a defined objective, which falls on harder targets and can erase the gains of scale

As these systems become more reliable the competitive advantage of offensive organizations shifts towards the cost of their compute, the efficiency of their orchestration and the operational knowledge their systems accumulate rather than the number of skilled operators they employ. That shift extends the industrialization of cyberattacks from technical execution into the organization of the criminal economy itself, where it reshapes how work is divided, which capabilities are bought as services and who can afford to take part.

Division of labor

Industrialization also changes who takes part in an industry, because it splits production into specialized suppliers and moves skill from the worker into the machine. Cybercrime had already done the first long before AI, with access brokers, malware developers, infrastructure providers and ransomware platforms each supplying one stage of an operation, and the packaging is now thorough enough that the Qilin ransomware service bundles tooling, infrastructure and negotiation support for affiliates who keep 80 to 85 percent of each ransom while a single rental service such as RedVDS, opens in a new tab supplied disposable servers to more than 10,000 users. AI does the second, since frameworks and distilled open models carry expertise that operators no longer need to have themselves, and like the factory worker who replaced the craftsman, a less skilled attacker can now run capabilities that once required an experienced team.

These two forces pull the industry in opposite directions. Deskilling lowers the cost of entry and brings in far more participants at the lower end, while the economies of scale in the attack factory reward whoever can spread fixed costs over the largest volume of attempts, concentrating the most capable operations in a few organizations that can afford the infrastructure, data and orchestration smaller actors cannot reproduce. The result looks less like a crowd of independent craftsmen and more like an industrial economy, with many small producers at the bottom and a few large ones at the top.

The new means of production

Industrialization has rarely eliminated skilled workers. It has moved them from doing the work to designing, running and improving the systems that do it, and offensive security is following the same path as operators spend less time on reconnaissance and exploit development and more on building workflows, validating results and deciding where the line should point.

When expertise becomes cheap and widely available it stops being the advantage, and scarcity moves to what cannot be downloaded, such as proprietary vulnerability data, realistic environments to test against, reliable execution infrastructure, access to valuable targets and the skill to orchestrate many agents efficiently. These are the inputs that will separate the few large operators from everyone else, and they are the same resources defenders now need to keep pace.

For decades offensive tools automated the technical tasks of an attack, and AI is now extending that automation into the reasoning, planning and adaptation that once required scarce human expertise.

Industrialization does not simply produce a faster craftsman. It changes the economics of production and, eventually, the structure of the entire industry.

Cyberattacks are approaching that transition.

Sources and further reading

Run the exploit factory before the attacker does

Apex turns the same industrial loop on your own code, web apps and APIs, proving what is exploitable and helping your team fix and retest it, so that every finding arrives as evidence rather than as noise.

Explore Apex Talk to our team

Products

Use Cases

AI SecOps

Application Security

The Industrialization of Cyberattacks

From human expertise to compute

Network discovery, analysis and scanning

Exploitation frameworks and web testing

Command and control and post-exploitation

Automated adversary emulation

AI-assisted research

Operational AI agents

The expertise that once capped offense can now be bought as compute

Within months, anyone can download frontier cyber capability and run it without safeguards

Mass production

Exploit factory

Control plane

Execution plane, parallel

Shared data plane, persistent

Evaluation plane, continuous

Division of labor

The new means of production

Sources and further reading

Run the exploit factory before the attacker does

Why Not Just Use Claude or Codex?

Cantina CEO Hari Mulackal on the Fable pull: the only clock you control Is how fast you fix

The Industrialization of Cyberattacks

From human expertise to compute

Network discovery, analysis and scanning

Exploitation frameworks and web testing

Command and control and post-exploitation

Automated adversary emulation

AI-assisted research

Operational AI agents

Mass production

Exploit factory

Control plane

Execution plane, parallel

Shared data plane, persistent

Evaluation plane, continuous

Division of labor

The new means of production

Sources and further reading

Run the exploit factory before the attacker does

Related articles

What we learned hunting for bugs in over 1500 codebases with AI

Why Not Just Use Claude or Codex?

Cantina CEO Hari Mulackal on the Fable pull: the only clock you control Is how fast you fix