Introduction
Artificial intelligence (AI) is transforming methods of cyberattacks. In times past, AI served as an auxiliary tool for tasks like crafting phishing messages and accelerating reconnaissance. However, there is a change, wherein AI systems are beginning to autonomously manage various stages of cyber operations with minimal human oversight. This prompts tough questions about the implications of AI operating beyond simply assistance.
Evidence of this transition emerged in July 2026 when the Taiwanese government agencies were reportedly targeted by an AI-enhanced cyber operation. As reported by the Financial Times, attackers utilized open-source AI agents that collaborated to map systems, identify security flaws, execute exploitations, and adapt their tactics in real-time based on findings. This operation involved multiple AI agents functioning simultaneously, rather than being employed as tools by human hackers (financialtimes.com). Taiwanese officials confirmed the cyberattack incorporated both conventional methods and AI-driven strategies (reuters.com). This incident shows the increasing integration of AI into offensive cyber operations and the need for investigation into the autonomy it exhibits.
The 2026 International AI Safety Report highlights that both criminal and state-affiliated actors are leveraging general-purpose AI for cyber operations. A study evaluating autonomous penetration capabilities indicates that, different AI models can independently discover and exploit vulnerabilities, albeit with varying success rates (Luo et al., 2026). Notably, the shift towards AI autonomy enables potential reconnaissance, target analysis, and decision-making without continuous human input, challenging traditional cyber operation paradigms (Adabara et al., 2025). This evolution necessitates a reevaluation of accountability and control in cybersecurity, as AI systems might lead to difficulties in attributing responsibility for actions taken autonomously.
The Taiwan Incident
A recent incident in Taiwan provides a good example of how AI may be changing offensive cyber operations. In July, Taiwanese government agencies were reportedly targeted in an overseas cyberattack that involved both conventional hacking techniques and AI agents. Taiwan’s Ministry of Digital Affairs confirmed that the attack was detected and contained, but did not publicly identify the attackers (reuters.com).
It attracted particular attention because cybersecurity researchers mention that attackers used multiple AI agents to perform tasks such as reconnaissance, identifying vulnerabilities, attempting intrusions, and adapting their strategies when an approach failed (financialtimes.com).
Taiwan has confirmed the use of AI-assisted techniques, but the claim that this was a completely autonomous, end-to-end cyberattack remains based largely on researchers’ analysis (reuters.com). The Financial Times described the activity as an unusually autonomous form of cyberattack, as AI agents were able to carry out several stages of the operation with limited human involvement.
The operation was still managed by human operators but is noteworthy because it shows a potential shift from AI helping hackers to AI carrying out increasingly intricate components of an aggressive cyber operation.
Human in the Loop & Human on the Loop
As AI’ s capabilities in cyber operations grow, the question of human control remains crucial. Two primary models exist; human in the loop and human on the loop. In the human in the loop model, AI can analyse information and recommend an action but a person must approve significant actions post analysis. This ensures direct oversight, particularly in high risk scenarios (International AI Safety Report, 2026). The International AI Safety Report 2026 advises this model for maintaining control in critical settings.
Conversely, the human on the loop model allows AI to operate independently within parameters set by humans, enabling quicker responses in cyber operations. This approach can be more practical for cyber operations because attacks and defensive responses can occur faster than humans can review every decision (International AI Safety Report, 2026).
While studies show AI’s evolving competence with minimal oversight, errors necessitate appropriate control measures. Requiring human approval for every step may slow an operation and reduce the advantage of AI’s speed. However, giving an AI agent too much independence could allow an error or unexpected decision to spread before a human can intervene (Moshin et al., 2025).
Sadovski et al., (2025) recommends matching the level of AI autonomy to the risk and complexity of the task, rather than treating human oversight as a simple yes or no choice. The 2026 cyber attack on Taiwan shows the importance of balancing human judgment and machine efficiency in cyber warfare. AI managed much of the operation & humans directed and made vital decisions. Future challenges will involve optimising human in the loop systems for control and speed against the risks posed by more autonomous AI systems.
Risks of AI Enabled Cyber Warfare
One important risk is speed and scale. AI systems can analyse large amounts of information, identify vulnerabilities, and support different stages of an attack much faster than human operators. The International AI Safety Report 2026 supports the assertion and reports that criminal groups and state-associated actors are actively using general-purpose AI in their operations.
AI can automate functions with less specialised expertise to conduct complex operations. This ability could increase the number of potential attackers and allow existing groups to conduct more operations simultaneously. RAND argues that agentic AI could reduce the human expertise constraint that has traditionally limited the scale of strategic cyber operations (rand.org).
Another significant risk posed is escalation in cyber operations, as they can operate at machine speed, outpacing human decision-making. This is especially alarming when targeting government systems or critical infrastructure. The recent incident in Taiwan exemplifies this concern, where AI agents executed multiple stages of an operation and adapted based on information acquired during the attack (financialtimes.com).
Concerns regarding attribution and accountability in cyberattacks have intensified with AI involvement. The challenge lies in pinpointing responsibility when AI systems autonomously select targets or alter methods, complicating responses from governments and adherence to current legal and military frameworks.
AI presents a dual-use problem; the same technologies can aid both attackers and defenders. While AI can enhance threat detection, vulnerability analysis, and attack response for defenders, there is the uncertainty regarding whether attackers or defenders will ultimately benefit more from AI advancements (International AI Safety Report, 2026).
AI enhances the effectiveness of cyberattacks, but the greater concern lies in its potential to alterate the dynamics between human decision making and machine action in future conflicts.
Conclusion
AI is not expected to fully replace human cyber operators soon, but the boundaries between human and machine led cyber operations are blurring. AI systems are automating stages such as reconnaissance and vulnerability discovery. The 2026 International AI Safety Report indicates that while completely autonomous cyberattacks are not yet realised, AI is significantly aiding both criminal and state sponsored cyber operations.
The recent attack on Taiwanese government systems involving AI agents illustrate this shift, from human led and machine assisted to human directed and machine executed operations (reuters.com). The challenge lies in maintaining meaningful human control as AI autonomy grows, which can enhance speed and scalability but also complicate accountability and error detection. Csernatoni, (2026) suggests these:
- The necessity for governance to regulate autonomous cyber operations
- Ensuring human oversight in high stake decisions
- At the same time preparing for a future partnership where humans define objectives and AI handles operational tasks.
The core concern remains; how much authority will be granted to AI in cyber operations, which could shape the future nature of cyber conflict.
References
Adabara, I., Sadiq, B. O., Shuaibu, A. N., Danjuma, Y. I., & Venkateswarlu, M. (2025). A review of agentic AI in cybersecurity: cognitive autonomy, ethical governance, and quantum-resilient defense. F1000Research, 14, 843.
Bengio,Y., Clare, S., Prunkl, C., Murray, M., Andriushchenko, M., Bucknall, B., Bommasani, R., Casper, S., Davidson, T., Douglas, R., Duvenaud, D., Fox, P., Gohar, U., Hadshar, R., Ho, A., Hu, T., Jones, C., Kapoor, S., Kasirzadeh, A., Manning, S.,… et al., “International AI Safety Report 2026” (DSIT 2026/001, 2026); https://internationalaisafetyreport.org.
Bengio,Y., Clare, S., Prunkl, C., Murray, M., Andriushchenko, M., Bucknall, B., Bommasani, R., Casper, S., Davidson, T., Douglas, R., Duvenaud, D., Fox, P., Gohar, U., Hadshar, R., Ho, A., Hu, T., Jones, C., Kapoor, S., Kasirzadeh, A., Manning, S.,… et al., “International AI Safety Report Executive Summary 2026” (DSIT 2026/001, 2026); https://internationalaisafetyreport.org/publication/2026-report-executive-summary
Blanchard, B., & Satter, R. Taiwan says it was targeted last month in AI-driven hacking campaign. reuters.com. August 23, 2026. (https://www.reuters.com/world/china/taiwan-says-it-was-targeted-last-month-ai-driven-hacking-campaign-2026-08-13/)
China-linked hackers hit Taiwan in unprecedented autonomous AI cyber attack . Financial times. August 23, 2026. https://www.ft.com/content/7d2ab3e0-9085-48f6-b38a-d90260d58795?syn-25a6b1a6=1 (https://www.ft.com/content/7d2ab3e0-9085-48f6-b38a-d90260d58795)
Csernatoni, R. (2026). When AI Agents Attack: Autonomous Cyber Operations and Europe’s Governance Gap.
Luo, J., Dai, J., Chen, Z., Xu, J., Wang, W., Duan, Y., … & Yang, M. (2026). The Emergence of Autonomous Penetration Capabilities in Large Language Model-Powered AI Systems. arXiv preprint arXiv:2606.13079.
Mohsin, A., Janicke, H., Ibrahim, A., Sarker, M. I., & Camtepe, S. (2025). A unified framework for human–ai collaboration in security operations centers with trusted autonomy. ACM Transactions on Internet Technology.
Sadovski, E., Aviv, I., & Hadar, I. (2025, September). Navigating the Human-oversight Dilemma in AI-based Systems. In 2025 IEEE 33rd International Requirements Engineering Conference Workshops (REW) (pp. 454-461). IEEE.
Sulmeyer, M. Artificial Intelligence and the Risk of Strategic Cyberattack . rand.org. August 13, 2026. https://www.rand.org/pubs/perspectives/PEA4901-1.html
