In 2025, the cybersecurity landscape is being reshaped by artificial intelligence. Models can help attackers produce convincing messages, write code, and research targets, while helping defenders sift through information at a scale no analyst could handle alone. AI is becoming a force multiplier on both sides. At the center of this transformation is the AI cybersecurity arms race.
The stakes are already high. Fortinet’s 2025 Global Threat Landscape Report recorded roughly 36,000 scans per second in 2024, a 16.7% increase in scanning activity over the previous year. Those detections measure reconnaissance, not successful attacks or AI’s contribution. They reveal the relentless pressure already facing defenders. AI-assisted targeting, deception, and code development threaten to add speed and reach to that existing machinery.
Governments are scrambling to keep up. A 2025 Ponemon Institute survey sponsored by MixMode found reported benefits in alert handling and analyst workload among surveyed U.S. organizations already using AI. China is building the computing infrastructure needed to develop and deploy models at scale. Meanwhile, criminals can draw on widely available AI tools without funding the research behind them. Governments must turn large investments into working defenses while their opponents look for the cheapest useful advantage.
This is no longer a linear contest between two superpowers. It is a three-way struggle between the United States, China, and a global network of threat actors. Each group is motivated by different incentives. National security, geopolitical dominance, and financial profit now all converge in this AI-driven battlefield.
As we move deeper into 2025, the AI cybersecurity arms race is reaching a critical inflection point. Generative AI is already being used for fraud and deception. Stolen accounts and vulnerable supply chains give attackers opportunities to exploit, and AI can reduce the effort needed to pursue them. Research into more autonomous tools points toward a further escalation. The window for proactive action is narrowing.
This blog examines seven key developments driving the AI cybersecurity arms race. From China’s stealth infiltration tactics to research on agentic malware and the policy gaps that leave nations exposed, we will explore how the landscape is evolving, where each side holds advantages, and where the greatest threats lie.
Despite billions in defense spending, it is not yet clear who is winning. What is increasingly clear, however, is that the rules of cyber warfare are being rewritten in real time.
Key Takeaways
- AI changes the effort required for cyber offense and defense. Assistance with language, code and research can make existing techniques faster and easier to scale.
- China's investment makes it a serious strategic competitor. Computing capacity and coordinated priorities give it resources to develop and deploy AI.
- U.S. innovation needs effective execution. Research strength must reach the people and systems responsible for protecting public infrastructure.
- Adaptive malware research deserves attention. RansomAI demonstrates a learning mechanism in a controlled setting, while human-led malicious use of generative AI is already documented.
- Lower barriers matter even without nation-state parity. An operator does not need elite capabilities to exploit a poorly protected account or carry out a convincing fraud.
- Stronger cooperation and enforcement are urgent. Existing law applies, but gaps in agreement, attribution and implementation can leave damaging activity difficult to contain.
The AI Cybersecurity Arms Race Explained
What defines the AI cybersecurity arms race in 2025
The AI cybersecurity arms race refers to the escalating global competition to develop and deploy AI for cyber offense and defense. Speed and scale are already at stake; greater autonomy could raise them further. In its January 2025 report on observed Gemini misuse, Google Threat Intelligence Group documented help with research, coding, troubleshooting, and content creation. The observed activity remained assistance to human-led work rather than a breakthrough in autonomous intrusion. Faster preparation still matters when an attacker can repeat it across many targets.
AI can compress work that previously consumed an operator’s time: drafting a message, explaining unfamiliar code, or troubleshooting a script. Defensive systems, in turn, use AI to triage alerts and flag anomalies across large volumes of data. Both sides are locked in a feedback loop of rapid iteration. An attacker who saves time on preparation can spend it on more attempts; a defender who cuts through alert noise can focus on the incident that matters.
What sets this competition apart is access. Nation-states are no longer the only actors able to draw on powerful AI models. The FBI’s December 2024 warning on AI-assisted financial fraud described criminals using generated text, cloned audio, and video impersonation to make deception more convincing. A criminal does not need to train a language model to use it for a persuasive lure. That lowers a practical barrier to entry and expands the reach of familiar scams.
Key players and their motivations
The primary actors in this arms race are the United States, China, Russia, North Korea, and decentralized hacking syndicates. Each brings unique resources and goals. For example:
- The United States invests in AI for defense and surveillance, relying on both public sector initiatives and private sector innovation.
- China treats domestic AI capacity as a strategic resource, with infrastructure investment and state priorities shaping its development.
- North Korea and Russia focus on disruption and revenue generation, often through cybercrime and cryptocurrency theft.
- Independent hacker groups seek profit, power, or disruption, enabled by AI-as-a-service and weak enforcement mechanisms.
Big tech companies, cybersecurity startups, and rogue developers also play critical roles. Tools developed for legitimate purposes are frequently repurposed and resold through illicit channels. The boundary between lawful innovation and malicious use continues to blur.
The infrastructure race: chips, compute, and data centers
Hardware remains a key differentiator. In June 2025, U.S. Commerce official Jeffrey Kessler estimated Huawei’s 2025 AI-chip production capacity at no more than 200,000 units. Meanwhile, a May report by SCSP and Strider identified more than 250 AI-focused data centers built or announced in China. Capacity estimates and announced facilities do not measure completed output, but the scale of the investment makes the strategic direction clear: China is pursuing the physical infrastructure needed to compete.
In contrast, U.S. firms hold major strengths in high-end GPUs, AI model development, and private-sector innovation. But export controls and global supply chain tensions complicate access and distribution of these resources.
Control of compute power is directly tied to the ability to generate, train, and deploy advanced AI models. As AI becomes more integral to cybersecurity strategy, nations are treating chip production and data infrastructure as critical to national defense.
Why 2025 marks a turning point
By 2025, the AI cybersecurity arms race has entered a new phase. Generative AI’s role in deception is documented, and technical assistance can make an operator’s existing skills go further. Research into agentic systems and reinforcement-learning malware adds the prospect of handing more decisions to software. The immediate danger is amplification; the longer-term danger is how much of the operation might be automated.
The result is pressure on the balance of power. Attackers can use assistance at selected stages without automating the entire operation. A faster research step or more convincing impersonation can still make a campaign more dangerous. Defenders must respond to that change while preparing for tools that may become more capable and adaptive.
As the next section explores, China’s state-directed approach puts coordination and scale at the center of this race. The United States faces a different challenge: turning its depth of private innovation into protection across a fragmented public sector.
China’s Strategic Use of AI in Cyber Warfare
China’s military doctrine and cyber centralization
China’s approach to the AI cybersecurity arms race places state priorities and infrastructure investment at the center of the competition. Central direction offers a potential advantage: funding, research, and deployment can be organized around national objectives. The strategic concern is the ability to concentrate resources on capabilities that serve both civilian and security goals.
Control over software vulnerabilities is part of that contest. A flaw kept secret can remain an entry point into someone else’s network; a flaw disclosed and fixed closes that opportunity. AI-assisted code analysis could accelerate the search for weaknesses. The stakes are therefore broader than who owns the fastest model: they include who discovers a weakness first and whether the affected systems are protected before an adversary acts.
Centralization may also shorten the path from research to operational use. A government able to align computing capacity, research priorities, and security institutions has an incentive to turn useful advances into capability quickly. That is the competitive challenge for the United States: a technical lead matters less if institutional delays prevent it from reaching the systems under threat.
Volt Typhoon and stealth infiltration tactics
Volt Typhoon shows the stakes of that competition. The Chinese state-sponsored group maintained persistent access to U.S. critical infrastructure, according to the February 2024 joint advisory from CISA, NSA, FBI, and partners. Its “living off the land” techniques used legitimate accounts and existing administrative tools to blend with ordinary network activity. Some access had persisted for at least five years.
The advisory did not establish AI as the mechanism behind those intrusions. Its warning was serious enough: the activity targeted communications, energy, transportation, and water and wastewater systems, with the assessed purpose of preparing for possible disruption. An intruder who looks like a legitimate administrator can be difficult to find long before AI enters the picture.
Now consider what AI assistance could add to such a problem. Faster research, help interpreting code, or more efficient troubleshooting could reduce the effort needed to navigate an unfamiliar environment. That is the amplification risk: adding useful assistance to methods that already threaten essential services. Critical infrastructure cannot wait for an entirely autonomous attacker before improving its defenses.
Regulatory positioning and global influence
China is not just investing in AI capacity. It is also shaping the rules for public use. Its Interim Measures for Generative AI Services, effective in August 2023, cover services offered to the public in China. They include labeling obligations for generated images and video, plus security-assessment and algorithm-filing requirements for services with public-opinion or social-mobilization capabilities.
Domestic regulation can give a government influence over how AI services reach the public. It does not resolve the separate question of how state-linked actors behave abroad. That tension matters for global governance: confidence in rules depends on conduct as well as the language of regulation.
Influence over AI standards and supply chains is as strategically important as software capability itself. Rules can shape access, constrain competitors, and give domestic institutions more control over an emerging technology. The battle for advantage extends beyond the code.
AI-for-hire: Covert and commercial capabilities
A less visible dimension of the arms race is the connection between commercial development and security capability. A model built for legitimate work can also assist reconnaissance, scripting, or deception. Governments and criminals can benefit from advances they did not invent. That makes the commercial AI ecosystem a strategic resource and a difficult target for oversight.
China’s access to useful computing resources therefore matters alongside its domestic development. Export controls and supply-chain restrictions seek to influence that access, while domestic investment seeks to reduce dependence. Both sides are trying to shape the conditions under which the next generation of AI capability can be built and deployed.
The contest is therefore about more than model quality. China’s state-directed investment challenges the United States to match its innovation with operational speed. A breakthrough that remains in a research environment offers little protection to an agency or utility facing an intrusion today.
The next section explores how the United States is responding to these developments and whether its strategy can keep pace with the rapidly evolving threat environment.
The U.S. Approach to AI-Powered Cyber Defense
AI integration in U.S. cyber infrastructure
The United States is integrating AI into security operations in an effort to stay competitive in the AI cybersecurity arms race. In the Ponemon/MixMode survey, practitioners at organizations already using AI reported benefits in alert handling and analyst workload. The practical goal is clear: help defenders find the important signal and act before an intruder can turn access into damage.
This includes using generative AI to summarize and interpret threat data, helping defenders prioritize incidents. Pattern-recognition systems can flag anomalies across large volumes of logs. The promise is a shorter path from warning to investigation and response. Delivering that benefit requires tools that work reliably in the environment they are meant to protect.
Despite these gains, the U.S. still faces significant challenges in scaling AI tools consistently across federal and critical infrastructure systems. Fragmentation remains a key issue, with varying capabilities and policies across agencies.
Public-private partnerships and defense contracts
To strengthen its position in the AI cybersecurity arms race, the U.S. has leaned heavily on public-private partnerships. Federal agencies such as DARPA and the Cybersecurity and Infrastructure Security Agency (CISA) work closely with major cybersecurity vendors and AI startups to develop next-generation defensive technologies.
Data-sharing and joint defense arrangements help connect that innovation with operational needs. For example, the Joint Cyber Defense Collaborative (JCDC) brings public and private actors together on cyber defense. AI-assisted detection is more useful when the right intelligence reaches a team with the authority and capacity to act.
Startups also play a vital role in piloting new approaches. Commercial research gives government access to ideas and tools it may not develop alone. This blending of innovation pipelines is a U.S. strength, but it only becomes a defensive advantage when useful tools reach the people and systems that need them.
The challenge, however, lies in operationalizing these technologies across a vast and often outdated IT ecosystem. Without uniform standards and consistent funding, breakthroughs often remain siloed rather than scaled.
Policy momentum: executive orders and standards
Policy is part of that response. The June 6, 2025 cybersecurity executive order amended the January order, directing access to cyber-defense research datasets and management of AI-software vulnerabilities and compromise. It also retained specific steps toward post-quantum readiness. The challenge is to turn those targeted requirements into working security practices across government.
Responsible AI use also demands reliable testing, auditing, and protection of civil liberties. An agency needs to know where a model fails, what decisions it influences, and who remains accountable. Moving quickly without that discipline can replace one form of exposure with another.
These moves are essential, given the global stakes of U.S. vs China cybersecurity. Policy alone cannot close the gap, but it can accelerate strategic alignment across government branches and industrial sectors.
Structural weaknesses and response delays
Despite its advantages, the United States faces structural issues that hinder its ability to fully respond to AI-driven cyber threats. The federal cybersecurity landscape remains fragmented, with overlapping jurisdictions and inconsistent readiness levels. In critical incidents, this lack of cohesion can delay response times and limit information sharing.
Moreover, the public sector often struggles to adopt and deploy advanced tools consistently. Outdated hardware, limited personnel training, and insufficient funding leave weaknesses that state-sponsored hackers and criminals can exploit. AI assistance can increase that pressure by helping an adversary work faster, while defenders remain constrained by systems and processes that are slow to change.
In the context of cyber warfare 2025, speed is critical. Slow procurement cycles and regulatory bottlenecks leave gaps that well-resourced adversaries can exploit.
As we’ll see in the next section, attackers do not need a better model than the government to gain an advantage. They need a useful tool, an exploitable weakness, and time before the defender responds. AI threatens to make that combination easier to assemble.
The Rise of Autonomous Hacking Tools
Generative AI’s role in evolving threat vectors
Generative AI is changing the nature of cyberattacks by reducing the work needed to prepare them. Models can help write a convincing message, adapt a script, or research a target. Human operators can use that assistance to move faster and repeat tasks at a scale that would otherwise require more people and time.
Phishing is a clear example. Generated text can imitate a professional tone and tailor a request to its recipient; cloned audio and synthetic video can reinforce an impersonation. The FBI's December 2024 warning describes criminals using these methods for financial fraud. They exploit trust, making the apparent familiarity of a voice or message a weaker safeguard.
The next frontier is greater autonomy. If a system can carry out a sequence of technical tasks and adjust to the results, the time between preparation and action could shrink further. That prospect makes agentic tools important to the AI cybersecurity arms race even while human-led misuse is the immediate, documented threat.
RansomAI and the future of self-improving malware
Reinforcement-learning malware shows where that next step could lead. The RansomAI research paper, submitted in June 2023, tested an agent that selected among six predefined encryption configurations on a Raspberry Pi testbed. It received feedback from an experimental behavioral anomaly detector using Isolation Forest.
The learning loop mattered: a configuration's effect on the detector influenced the agent's next choice. This was a controlled research demonstration, not a report of criminals deploying ransomware that learns across victims. Its significance lies in the mechanism—using feedback to seek encryption behavior that attracts less attention.
For defenders, the warning is concrete. Detection systems are part of the environment an adaptive attacker could learn to work around. Testing a defense against one fixed behavior may leave it poorly prepared for software that changes its choices. RansomAI turns that concern into a specific problem for security researchers to investigate.
Credential theft at unprecedented scale
Credential-based attacks already have an enormous supply of exposed information to draw on. Fortinet's report describes 1.7 billion stolen credential records shared in underground forums in 2024. Separately, the broader volume of compromised records shared there rose 42% from the previous year. Those figures count records, not attacks or unique victims.
Automation can test stolen logins against other services; AI adds assistance with the surrounding work, from adapting scripts to producing convincing follow-up messages. A password leak, a plausible impersonation and an account with excessive privileges can become parts of the same operation. Faster preparation increases the pressure on defenders to detect suspicious access and limit what follows it.
The speed of that response matters. Even a short delay can give an intruder time to reach more systems or accounts. Automated analysis can help defenders find relevant patterns across large datasets, but it needs to be paired with access controls and containment. A useful alert must lead to action.
Criminal services and the spread of AI capability
Criminal services spread tools and expertise beyond the people who created them. Generative AI extends that pattern: an operator can obtain help with language, code and content production without building a model or employing every specialist those tasks once required.
This changes the economics of an attempt. Better wording can make a fraud more credible; faster translation can extend its reach; coding assistance can reduce the work involved in adapting a tool. The assistance need not complete an entire attack to be valuable to someone directing it.
Access alone does not eliminate the need for skill, infrastructure or a vulnerable target. It can nevertheless remove a bottleneck. The AI cybersecurity arms race is therefore also a contest over how widely useful capabilities spread and how quickly defenders close the opportunities those capabilities exploit.
The next section examines whether institutions can respond at the pace that this wider access demands.
Are Hackers Outpacing Governments?
Agentic AI and the rise of independent threat actors
Agentic AI is a particularly important development because it aims to connect individual tasks into longer workflows. A system that plans actions, uses tools and responds to results could reduce the amount of step-by-step direction needed from an operator. Applied maliciously, that could shorten the path from identifying an opportunity to acting on it.
Governments typically operate through layered approvals, legacy infrastructure and strict protocols. Attackers can often change their approach without those institutional handoffs. Adding useful automation to that asymmetry raises the pressure on public agencies to make their own detection and response processes faster.
The immediate lesson is not to wait for a fully independent attacker. Google's January 2025 observations already showed assistance with human-led research, coding and content creation. Institutions need to address those gains now while preparing for the greater autonomy that agentic development could bring.
Cybercrime-as-a-service lowers the barrier to entry
Cybercrime-as-a-service lets an operator use tools or infrastructure developed by someone else. AI assistance can reduce additional barriers, especially in tasks involving language and code. The result is a wider pool of people able to attempt parts of an operation without first acquiring every underlying skill.
A novice does not become a state-sponsored team by generating a script or a message. But defenders cannot assess the risk only by asking whether the newcomer has elite capabilities. A less skilled operator who produces a credible fraud, reaches a poorly protected account or reuses an effective technique can still cause substantial harm.
The important question is where assistance removes effort that previously limited an attacker. Governments and industry need to identify those points, share useful findings and interrupt the resulting workflows. Waiting for proof of nation-state parity would set the threshold for action far too high.
China’s speed advantage vs. U.S. innovation depth
The AI cybersecurity arms race between the U.S. and China reveals a critical tension: depth of innovation versus the ability to put it to work. China's centralized priorities can concentrate resources on a strategic objective. The United States draws strength from private-sector research and model development, but integrating those advances across public institutions is a separate challenge.
The computing investments discussed earlier make this more than a contest of ideas. They create capacity to develop and deploy AI at scale. Turning that capacity into reliable operational tools still requires integration, expertise and effective decisions about their use.
The comparative advantage may therefore depend as much on who deploys technology effectively as on who develops it first. For U.S. agencies, reducing friction between research, procurement and operational use belongs alongside funding the next breakthrough.
Defensive lag and policy paralysis
Slow procurement, inconsistent coordination and outdated infrastructure can give attackers time that defenders cannot afford to lose. A useful detection model does little good if its output waits for another team, an unclear authority or an incompatible system before anyone can act.
AI increases the urgency because it can accelerate malicious research, scripting and deception. Public institutions need both better analysis and the capacity to respond to it. Otherwise, an improvement in attacker productivity may reach real operations sooner than a defensive improvement reaches the systems it is meant to protect.
This is where the warning that hackers could gain ground is strongest. It concerns the gap between available technology and operational readiness. Reducing that gap is a practical defense against both existing techniques and the more capable tools now being developed.
The consequences of slow or ineffective protection are already visible in major breaches. The following examples show the stakes, alongside direct evidence of malicious AI use.
Major Breaches and Documented AI Misuse
Lazarus Group and the $1.4 billion crypto theft
Few incidents illustrate the financial stakes more clearly than the Bybit cryptocurrency theft. The FBI attributed the February 21, 2025 theft to North Korea, describing approximately $1.5 billion in stolen virtual assets and identifying the activity as TraderTraitor.
Bybit's preliminary forensic account described compromised Safe developer credentials, altered wallet infrastructure and deception of transaction signers. Elliptic traced the subsequent money movements through asset conversions, wallets and cross-chain services. These accounts establish a high-stakes attack on a trusted workflow, without establishing an AI-directed breach or machine-learning laundering.
That trust problem is central to the wider AI risk. Faster technical assistance and more convincing impersonation could help other attackers prepare operations against people authorized to approve valuable transactions. The task for financial institutions is to strengthen the trusted workflows that such assistance could help an adversary target. For a related case involving deceptive recruitment, see North Korean Deepfake Job Scam: 7 Shocking Red Flags.
ChatGPT exploited for disinformation and surveillance
OpenAI's June 5, 2025 report provides direct evidence of malicious AI assistance. Four of its ten case studies were assessed as likely originating in China. The report describes influence activity, social engineering and technical work involving generated content, coding and scripting assistance.
These activities show why generative AI matters beyond intrusion itself. Producing deceptive narratives in multiple languages or helping with technical work can reduce the workload required to sustain an operation. A human-led campaign can become easier to repeat or extend without an autonomous system running it from start to finish.
The cases do not all establish state direction or the eventual impact of the assistance. They do establish that malicious actors were using the tools. For policymakers, information operations and network security therefore belong in the same discussion of how AI changes the resources an adversary needs.
The risk of misconfigured AI defenses
AI can also fail defenders. A poorly tuned model may miss relevant behavior or flood analysts with low-value alerts. Overreliance on its output can then turn a promised efficiency gain into another source of blind spots.
This is why deployment requires oversight, suitable data and testing in the environment the system will protect. Defensive AI must support a process that can identify a problem and act on it. An institution that automates analysis without improving response risks acquiring a false sense of security rather than a stronger defense.
Real-world impact on civilian infrastructure
The consequences of cyber disruption are not confined to information on a screen. Hospitals, utilities and public services depend on access to functioning systems. Disruption can interrupt care and essential operations, putting safety and trust at risk.
As detailed in our analysis of hospital vulnerabilities, service continuity belongs at the center of cyber planning. If AI makes malicious preparation and deception easier, institutions already managing fragile dependencies may face more pressure. That is a reason to strengthen recovery and coordination alongside prevention.
These risks expose the gap between technological promise and operational readiness. The AI cybersecurity arms race is not just about who has better tools. It is about who can deploy them responsibly and effectively while keeping essential services running.
The global implications demand attention. The next section explores whether international cooperation can keep pace with an escalating digital arms race.
Global Implications: Cyber Arms Treaties or Chaos?
The risk of AI-driven global destabilization
As the AI cybersecurity arms race accelerates, gaps in cooperation and enforcement become more dangerous. Lower-cost assistance with deception and technical work can put useful capabilities within reach of more actors. The resulting pressure extends beyond the two largest AI powers.
Cyber operations can disrupt services across borders. If AI helps operators prepare more attempts or adapt their work faster, healthcare, transportation and energy systems may face a more demanding threat environment. The policy danger is that capability spreads faster than institutions build the safeguards and working relationships needed to contain it.
Attribution remains difficult. Misinterpretations or false flags could lead to retaliation based on incomplete information. Better shared expectations, investigative cooperation and credible consequences are therefore security measures, not merely diplomatic aspirations.
Existing rules and gaps in AI-specific agreement
As of June 2025, there was no comprehensive treaty written specifically for AI-enabled cyber warfare. Existing law still mattered: the 2021 U.N. working-group report on state use of information and communications technologies reaffirmed that international law, particularly the U.N. Charter, applies to state ICT activity. States nevertheless disagreed on important questions of application.
Tallinn Manual 2.0 offers nonbinding expert analysis of how international law applies to cyber operations. The Budapest Convention on Cybercrime creates obligations concerning offences, investigative powers and international cooperation. Neither is a dedicated AI arms-control regime, and enforcement still depends on jurisdiction, implementation and cooperation.
The gap is therefore practical as well as legal: identifying responsibility, applying existing obligations and securing agreement where more specific rules are needed. States and private actors are not immune from consequences, but inconsistent enforcement can leave room for damaging activity. Stronger governance can build on existing obligations by making their application and enforcement more credible.
China’s domestic regulation vs. global engagement
China's domestic AI rules and its international security conduct raise a difficult question of trust. The Interim Measures discussed earlier regulate generative-AI services offered to the public in China. Their labeling and assessment requirements do not settle how the state or other actors will behave internationally.
OpenAI's report documents likely China-origin misuse, while the Volt Typhoon advisory describes a separate state-linked intrusion campaign. Together they give other governments serious reasons to scrutinize conduct, without merging the two into a single proved AI-directed operation. Domestic regulation is only one part of an assessment of responsible behavior.
In the broader U.S.–China rivalry, competing security priorities can make agreement difficult even where all sides have an interest in stability. Credible commitments will require observable behavior and accountability. Related questions about trusted identity infrastructure arise at home; our analysis of the Trump digital ID repeal considers that separate policy debate.
What a meaningful treaty would require
Creating an effective cyber arms agreement in the age of AI would require several key components. First, nations would need to agree on verification protocols. These could include independent audits, AI model classification systems, and mandatory reporting of high-risk capabilities.
Second, attribution mechanisms must improve. Without credible methods for identifying those responsible for an operation, enforcement becomes much harder. Third, enforcement must carry real consequences consistent with applicable law. Options could include sanctions or agreed restrictions for violators.
Finally, major players must participate. A treaty without the United States, China, or the European Union would lack legitimacy and reach. Success will depend on aligning these actors around shared interests in stability, even if their strategic goals diverge.
The next section evaluates whether any party is truly winning the AI cybersecurity arms race, or if we are heading toward a future defined by persistent vulnerability and decentralized conflict.
Who’s Really Winning and What Happens Next?
Comparative advantage: U.S., China, and hacker groups
As the AI cybersecurity arms race unfolds, the question of who holds the advantage is becoming more complex. Rather than a binary contest between two nations, the landscape now involves a multi-front battle among the United States, China, and decentralized hacker ecosystems.
The United States has major strengths in foundational AI research, advanced chip design and private-sector innovation. It has world-class universities, dominant cloud providers and a vibrant cybersecurity industry. Its challenge is translating those resources into protection across a fragmented public-sector environment.
China's investment and centralized priorities offer a different route to capacity. Concentrated state direction can sustain large programs, while commercial research can explore a wider range of approaches. In either system, deployment, reliability and coordination determine which advances become useful tools.
Hacker groups operate across the gaps between these institutions. They can adopt a useful technique without waiting for a government-wide program. AI assistance increases the importance of that agility: a gain in writing, scripting or research can matter as soon as an operator can use it. Defenders need to shorten their own route from a useful discovery to action.
Most urgent trends to monitor in 2025–2026
Several trends stand out as critical to the direction of the AI cybersecurity arms race:
- Agentic malware research: Whether controlled adaptive techniques become reliable operational tools is a critical development for defenders to monitor.
- LLM misuse: Assistance with phishing messages, deceptive content, code and research can make human-led campaigns easier to sustain and expand.
- Synthetic impersonation: Cloned voices and generated video challenge verification based on familiar appearance or sound.
- AI and supply chains: Trusted software and service relationships remain valuable targets. AI assistance could make preparation against those relationships faster or easier.
These trends put a premium on defenses that can adapt and on institutions that can act quickly. Established controls remain essential; they need to be supported by analysis and response processes able to keep pace with changing threats.
What governments must do to close the gap
Closing the gap will require action on multiple fronts. Governments must invest in AI-specific cyber infrastructure, including faster data pipelines, threat intelligence platforms, and real-time analytics. Funding should be allocated not just to new tools, but to deployment and integration across agencies.
International cooperation is also key. The AI cybersecurity arms race is global. No single nation can defend against every threat alone. Governments must expand data-sharing agreements, co-develop detection models, and coordinate response plans with both allies and industry partners.
Streamlining policy is equally important. Agencies need the legal authority and technical capacity to respond rapidly to evolving AI-driven cyber threats. This includes faster procurement pathways, more flexible hiring processes, and clear mandates for AI integration.
Conclusion: Is anyone truly in control?
No actor appears to have full control of the battlefield. The United States has formidable technological resources, China is investing heavily in capacity, and hacker groups exploit weaknesses across both national and institutional boundaries. The contest turns on how effectively each can put its capabilities to work.
AI raises the pressure by reducing the effort needed for important parts of an attack. It can help a human operator prepare more convincing deception, work through technical problems and repeat a process. Research on adaptive malware points toward further change. If those gains reach attackers faster than institutions improve protection and response, the advantage can shift toward offense.
The AI cybersecurity arms race is not just a contest of resources. It is a test of agility, coordination and long-term vision. Governments should treat deployment and response capacity as seriously as the next technical breakthrough. We cover emerging trends like this in our newsletter, with weekly breakdowns of the tools, threats and policies shaping cybersecurity and AI. Subscribe here to follow that work.
FAQ Section
Q1: What is the AI cybersecurity arms race?
It is the competition among governments, companies and malicious actors to use AI for cyber offense and defense. AI assistance with language, code and analysis changes the effort required for important tasks. The race concerns how quickly those capabilities can be deployed, alongside research into more autonomous systems.
Q2: Why is Volt Typhoon a big deal?
The joint advisory describes Chinese state-sponsored actors maintaining access to U.S. critical infrastructure through legitimate accounts and system tools. Such access could support future disruption. The documented campaign shows how difficult persistent intrusion can be to detect; it is not a confirmed example of AI-directed hacking.
Q3: Can AI defend against AI-based attacks?
Yes. AI can assist with threat analysis, prioritization and investigation, whether or not the attacker used AI. Success depends on suitable data, testing and a response process able to act on the results. Deployment and coordination are therefore as important as model capability.
Q4: What are agentic AI threats?
Agentic systems can plan actions, use tools and respond to results with less step-by-step direction. Malicious use could extend automation across more of an operation, reducing the time an attacker needs to act. Reliability and access determine how far that possibility can become an operational capability.
Q5: Are there any international rules governing AI in cyberwarfare?
Existing international law can apply to state cyber operations, and cybercrime obligations can cover relevant conduct involving AI. As of June 2025, there was no comprehensive treaty written specifically for AI-enabled cyber warfare. More specific agreement and effective enforcement remain important gaps in the international response.