AI intensifying cybersecurity risks

Last month, a hacker used a prompt of fewer than 100 Chinese characters in directing artificial intelligence (AI) agents to attempt to break into more than 1,200 Web sites, ultimately succeeding in more than 100 cases. Norwegian cybersecurity firm NorForge discovered the hacker by scanning for activity linked to Cairn, an open-source AI problem-solving platform popular among hackers. The agents were found autonomously probing for security holes and testing infiltration strategies in an operation that is estimated to have cost the hacker just US$50 per successful hack. Victims included several high-profile US corporations, including farm equipment producer John Deere, home appliance maker SharkNinja and engine manufacturer Cummins. The incident highlights how much of the work that once demanded long hours on the part of hackers can now be offloaded to autonomous agents.

Most alarming is not the volume of attacks, but the agents’ capacity to autonomously improve and adjust their strategies. In one instance detected by Norforge, AI agents initially failed to break into a company’s public-facing application, but subsequently scoured the databases of associated infrastructure and found a publicly accessible control panel. After hours of trial and error, they were able to leverage a weakness in data processing methods to trick the server into issuing malicious commands. These were not attacks running directly off prepared scripts; they were agents continuously exploring and adapting to vulnerabilities in the environment.

The economics of cybersecurity is undergoing a radical shift. Traditional hackers had to invest significant time and energy to search for weak points, adjust their code and trial attack pathways, but AI agents can do this work both faster and without rest. All attackers must now do is pay modest user fees, and sit back as the agents do the rest. While firms and Web sites must safeguard all possible entry points, AI agents need only to find a single way in. With attack volumes surging, companies might struggle to maintain their defenses even with increased information security budgets and personnel.

Around the same time that NorForge released its findings on Oct. 19, the US Department of Homeland Security ordered that China’s DeepSeek products and related services be removed from all US civilian agency information technology systems, reportedly due to data security concerns and supply chain risks. However, this particular attack made use of both DeepSeek and Anthropic’s Claude Code, demonstrating that AI risks are by no means specific to a model’s country of origin.

While there might be a case to be made for restricting the use of certain high-risk models in government, criminal use can be made of AI systems from any country. For practical uses of AI agents, it should also first be established what permissions have been granted, whether they are subject to monitoring and whether they can be halted in the event of anomalous behavior.

Taiwan’s semiconductor fabs, data centers, power grid and financial systems are all highly dependent on digital networks, and AI is increasingly involved in their maintenance and process management. This case of Web site break-ins shows us that equipment suppliers, remote maintenance accounts and outsourced service platforms can all provide attackers with entry points. Because semiconductor supply chains extensively involve third-party contractors and equipment vendors, even the tightest cybersecurity measures for chip fabs cannot prevent the rise of overall risks presented by security breaches elsewhere.

It is also worth noting that just as the NorForge-discovered hacker was funding their entire operation with just US$5,000, Anthropic was agreeing to a US$45 billion deal to secure more computing power. With the pursuit of ever more powerful models being backed by tens of millions of dollars from their developers, bad actors need only spend a pittance in comparison to leverage a piece of this computing power and pose serious security threats.

With AI capabilities appearing to improve faster than security frameworks can adapt, the steepest bill to foot might not be information security investments, but the shutdowns, losses and erosion of trust resulting from cyberattacks.

Taiwan is an important global supplier of AI chips and servers, but hardware specialization does not make it less vulnerable. While building chips that make AI run even faster, it cannot allow protections for the industrial systems that underpin their manufacturing to become a weak link. As AI recalibrates the cost to hackers for launching digital offensives, governments and industry must also re-evaluate the price they are willing to pay for protection.

Liao Ming-hui is an assistant research fellow at the Chung-Hua Institution for Economic Research.