The most important change in cyberattacks is not that hackers suddenly became superhuman. It is that artificial intelligence is allowing relatively small teams to perform more of the work that once required large groups of experienced operators.
Recent reporting describes a suspected China-nexus campaign that used an AI-enabled penetration-testing framework against advanced semiconductor companies, defence organisations and civil-society groups. Palo Alto Networks researchers said the attackers targeted at least 19 organisations and gained access to, or exfiltrated data from, 10 of them. The campaign reportedly obtained valid credentials, created administrator accounts and exposed more than one million personal records at one victim organisation.
The case is significant because it involves industries at the centre of national security and economic competition. Semiconductor firms hold intellectual property, manufacturing information and supply-chain secrets. Defence companies hold technical data, contracts and research connected to weapons and military systems. A successful intrusion into either sector can have consequences far beyond the theft of ordinary corporate files.
But the story needs careful wording. AI did not independently decide to attack the companies. Human operators selected targets and built the framework. AI helped automate and accelerate the operation. That distinction matters because it shows both the danger and the remaining opportunities for defence.
What happened in the semiconductor and defence campaign?
According to Cybersecurity Dive’s account of Palo Alto Networks’ Unit 42 research, the attackers used a framework called Cairn together with Chinese open-source models DeepSeek and Zhipu GLM. Researchers tracked the activity under the designation CL-STA-1238 and assessed it as part of a China-nexus intelligence-collection effort.
The campaign reportedly focused on at least 19 organisations in sensitive sectors, including advanced semiconductors and the US defence industry. Ten of the targeted organisations were reported to have suffered successful access or data theft.
The reported results show why the campaign matters:
•The attackers obtained dozens of valid credentials.
•They created administrator accounts inside a production system.
•They exposed more than one million personal records at one organisation.
•They used automation to work across more targets than a conventional hands-on-keyboard team could efficiently manage.
The exact division of labour between humans, the Cairn framework and the AI models is not fully public. That uncertainty should not be ignored. Security companies often publish early findings while investigations are still developing, and attribution claims can change as more evidence becomes available.
Still, the pattern is clear enough to be consequential: AI-assisted tooling is being used to increase the speed and scale of real intrusion campaigns against high-value organisations.
Why chip companies are especially attractive targets
Semiconductor companies sit at the intersection of technology, defence, telecommunications, artificial intelligence and national economic policy.
A chip manufacturer’s most valuable information may not be stored in a single obvious database. It can be spread across design systems, engineering repositories, fabrication documentation, supplier networks, test environments and employee accounts. An attacker who obtains legitimate credentials may be able to move through several connected systems while appearing to be an authorised user.
The target is not always a finished chip design. Intelligence can include manufacturing processes, production yields, equipment configurations, packaging methods, supply-chain information, customer requirements and research schedules. Even partial access may help a competitor or state-linked organisation understand where a company is ahead, where it is vulnerable or what it plans to build next.
The industry also depends on a broad ecosystem of contractors, software suppliers, equipment makers, cloud services and logistics providers. That interconnectedness creates efficiency for business, but it creates more possible entry points for attackers.
AI makes that complexity more useful to the attacker. A model can help organise publicly available information, interpret technical documentation, compare exposed services and summarise the relationships between people, systems and suppliers. The human attacker still needs judgment, but the amount of preparation that can be delegated has increased.
Why defence companies face an even higher risk
Defence companies are attractive because their networks often contain information with strategic value. A breach may expose technical specifications, research programmes, procurement data, subcontractor relationships or information about military capabilities.
The damage is not limited to the company that was directly hacked. Defence production depends on thousands of suppliers, including small specialist firms that may not have the same security budget as a major contractor. A smaller supplier may hold a valuable component design or maintenance document even if it is not itself a famous target.
Defence organisations also face a difficult balance. They must allow engineers, researchers and suppliers to collaborate across complex environments, but every connection adds identity, access and monitoring challenges. Strict isolation can protect sensitive systems, yet excessive isolation can slow production and create workarounds that introduce new risks.
AI-enabled attacks are dangerous in this environment because they can search for weak links faster than a human team. An attacker does not need to defeat the entire defence ecosystem. It may only need to find one over-privileged account, one forgotten internet-facing service or one supplier whose credentials open a path into a more valuable environment.
What AI is actually doing for attackers
The most useful way to understand AI in cybercrime is as a force multiplier. It can reduce the time needed for tasks that are repetitive, language-heavy or dependent on searching large amounts of information.
In a sophisticated campaign, AI may help with:
•Mapping the target’s digital infrastructure.
•Summarising technical documentation and public records.
•Identifying potentially weak services or configurations.
•Generating or adapting code for testing.
•Sorting stolen credentials and data by likely value.
•Creating convincing messages or impersonation attempts.
•Recording what has already been examined and suggesting the next step.
These tasks may look ordinary when considered separately. Their significance comes from the way they can be linked into a longer workflow.
Anthropic’s account of an earlier suspected Chinese state-sponsored campaign illustrates this agentic model. The company said attackers used its Claude Code tool to attempt infiltration of roughly 30 global targets, including technology companies, financial institutions, chemical manufacturers and government agencies. Anthropic assessed that AI performed 80% to 90% of the campaign, while human operators intervened at a small number of critical decision points.
Anthropic also acknowledged that the model made mistakes. It sometimes hallucinated credentials or claimed to have extracted information that was actually public. This is an important limitation. Automation can increase speed, but it can also generate false confidence and waste effort.
The broader lesson is that attackers do not need AI to be perfect. If the cost of trying an attack falls enough, an imperfect automated process may still be valuable.
The threat is moving from assistance toward agency
Earlier cyber tools generally helped people perform individual tasks. A scanner found exposed services. A script tested a configuration. A stolen-credential tool attempted logins. A human operator decided what the results meant.
Agentic systems can connect more of those steps. They can observe a result, choose a follow-up action, call another tool and continue working with limited supervision.
Google Threat Intelligence Group reported that adversaries were experimenting with this kind of automation in several areas. Its 2026 tracker described a suspected AI-assisted zero-day development effort, AI-augmented malware and malware capable of using an AI model to interpret a victim device and generate commands.
Google also reported a cybercrime campaign involving a zero-day vulnerability that researchers believed had been developed with AI assistance. The company said the vulnerability allowed a two-factor authentication bypass but still required valid credentials. That qualification is important: AI-assisted vulnerability discovery does not mean every attack can defeat every security control.
The same report described PROMPTSPY, an Android backdoor whose AI component was designed to interpret a device’s visible interface and guide follow-on interactions. This is a different type of threat from a semiconductor espionage campaign, but it shows the same direction of travel: models are being connected to tools and environments so that they can act, not merely produce text.
How fast is the wider threat landscape moving?
CrowdStrike’s 2026 Global Threat Report said attacks by AI-enabled adversaries increased by 89% in its data and that 82% of detections in 2025 were malware-free. It also reported that the average eCrime breakout time had fallen to 29 minutes, with the fastest recorded case taking only 27 seconds.
These figures come from CrowdStrike’s proprietary threat intelligence, so they should not be treated as a universal measurement of every cyberattack. They are nevertheless useful indicators of the operational problem: defenders may have less time between an initial compromise and an attacker moving deeper into a network.
Microsoft’s 2026 Digital Defense Report describes AI as changing the “physics” of cybersecurity by compressing attack timelines, lowering the cost of sophisticated capability and enabling greater speed, scale and autonomy. Microsoft also stresses that attackers are targeting AI systems themselves, while legitimate AI tools can be abused to generate malicious commands or steal sensitive data.
This creates a two-sided race. Defenders can use AI to analyse logs, prioritise vulnerabilities and identify unusual behaviour. Attackers can use AI to generate more attempts, adapt their methods and exploit gaps between identity, cloud, endpoint and network monitoring.
Why identity has become the central battlefield
Many modern attacks do not look like a dramatic break-in. They look like legitimate users signing in, accessing a cloud service or using an application with permissions they already possess.
That is why identity is so important. Microsoft identifies identity as the primary control plane for defence and recommends phishing-resistant multifactor authentication, passkeys, disciplined identity governance, tiered administration and strong privileged-access controls.
The semiconductor and defence incidents described in recent reporting reinforce this point. Obtaining valid credentials can be more valuable than deploying noisy malware. Once attackers appear to be authorised users, security teams must distinguish normal business activity from malicious use of legitimate access.
AI can help attackers sort through identities and permissions at scale. It can also help defenders identify abnormal combinations, such as an account accessing an unusual system, creating a new administrator or downloading data outside its normal working pattern.
The goal is not to treat every unusual action as proof of compromise. That would overwhelm security teams with false alarms. The goal is to connect identity, endpoint, cloud, application and network signals quickly enough to recognise a campaign while it is still developing.
Are hackers now fully autonomous?
No—not in the way the most dramatic headlines sometimes suggest.
The recent cases show increasing automation, but human operators still matter. They select strategic targets, provide infrastructure, choose objectives and decide when an operation is worth continuing. AI systems also make errors, misunderstand context and can be blocked by missing permissions or unfamiliar environments.
Even Anthropic’s description of an 80%–90% AI-performed campaign came from the company’s own investigation into activity involving its tool. It is an important case study, but it should not be read as proof that all cyberattacks are now mostly autonomous.
The more defensible conclusion is that the amount of human effort required for some campaigns is falling. A small team can now attempt work that previously required more operators, more time and more specialised expertise.
That is dangerous even if the AI still needs supervision.
Why the semiconductor and defence incidents are a warning
The combination of target and technique is what makes these cases alarming.
A campaign against an ordinary consumer website might cause financial loss or privacy harm. A campaign against a chip designer or defence supplier can affect strategic competition, weapons development, industrial capacity and national resilience.
At the same time, AI-assisted attack frameworks can be reused. Once an attacker has developed a system that can identify exposed services, test common weaknesses, process credentials and organise stolen data, the same framework may be redirected at many organisations.
This changes the economics of cyberespionage. The attacker may not need to make every campaign highly customised from the beginning. It can automate broad discovery first and devote human attention only to the most promising targets.
That is why a campaign involving 19 organisations and 10 reported compromises matters even if it did not produce a catastrophic disruption. It demonstrates an operational model that can be scaled.
What should companies do differently?
The answer is not simply to buy an AI security product. Organisations need to reduce the opportunities that automated systems can exploit and improve the speed at which they detect misuse.
The first priority is identity. Companies should remove unnecessary privileges, protect administrator accounts, require phishing-resistant authentication for sensitive access and monitor the use of service accounts, application identities and automated agents.
The second is exposure management. Internet-facing systems, edge devices, remote access tools, cloud workloads and supplier connections need continuous review. A vulnerability that is technically known but not visible to the security team is still an active risk.
The third is segmentation. Sensitive chip designs, defence data and production systems should not be reachable through a single broadly trusted account or flat internal network. Segmentation cannot eliminate compromise, but it can limit how far an attacker can move.
The fourth is data control. Organisations need to know what sensitive information exists, where it is stored, who can access it and whether unusual downloads or transfers are being detected. Data classification is useful only when access rules and monitoring enforce it.
The fifth is incident response. Security teams should practise scenarios in which credentials are stolen, administrator accounts are created and attackers move through cloud and production systems without deploying obvious malware.
Finally, organisations should govern their own AI tools. Microsoft warns that attackers can target AI systems as well as use them as tools. Companies need controls for what internal models can access, which actions agents can take and when human approval is required.
The defence side of the AI race
The same technology that helps attackers can improve security operations.
AI can help defenders examine large volumes of telemetry, correlate weak signals, summarise incidents, prioritise vulnerabilities and suggest containment steps. It can help analysts understand unfamiliar code and compare current activity with known threat patterns.
But defensive AI must be deployed carefully. A system that can read sensitive logs may become a target itself. An agent that can change firewall rules, disable accounts or isolate production systems needs strong permissions, testing and human oversight.
The principle should be controlled acceleration. AI should make defenders faster without giving an untested system unlimited authority over critical infrastructure.
Threat intelligence sharing also matters. An attack framework used against one semiconductor company may be detected earlier at another if indicators, tactics and behaviour patterns are shared quickly. Companies that treat incidents as private reputational problems can unintentionally protect attackers from their next target.
Is cybersecurity entering a dangerous new era?
Yes, but the danger is more practical than cinematic.
The immediate risk is not that machines have replaced every human hacker. It is that AI is lowering the cost of reconnaissance, increasing the number of targets that can be tested and shortening the time defenders have to respond. It is also making it easier for attackers to operate across complex environments involving identities, clouds, suppliers and edge devices.
The most exposed organisations are not necessarily those with the largest servers. They are organisations whose valuable information is connected to many systems and partners, whose privileges are difficult to control and whose security teams cannot correlate signals quickly.
Chip and defence companies are especially important because their information can shape the future of entire industries and national-security programmes. A successful intrusion may not be visible to the public for years, yet its effects can appear in lost technological advantage, altered supply chains or accelerated military development.
The right response is neither panic nor denial. It is to assume that attackers will use AI to move faster, then design security controls around that assumption.
Hackers are using AI
to attack sensitive organisations, including semiconductor and defence companies. Recent reporting describes a suspected China-nexus campaign that used AI-enabled penetration-testing tools against at least 19 organisations and reportedly succeeded against 10.
The technology is also being used to automate reconnaissance, vulnerability research, credential handling, data sorting and malware operations. Anthropic, Google, CrowdStrike and Microsoft all describe a broader transition toward faster, more scalable and more autonomous cyber activity.
But the evidence does not show that human hackers have become irrelevant. Humans still choose targets, construct campaigns and make strategic decisions, while AI systems remain imperfect and can produce false results.
That is precisely why the new era is dangerous. Attackers do not need perfect artificial intelligence. They only need tools that allow a small number of people to attempt more intrusions, against more targets, at greater speed.
For the companies building the chips and defence systems on which modern societies depend, cybersecurity can no longer be treated as a narrow IT function. It is now part of industrial security, national resilience and strategic competition.
