Smarter Machines, Smarter Attacks: Understanding Cybersecurity Threats in the Age of AI

By Bhupendra Thapa

"As artificial intelligence reshapes the future, cybersecurity becomes the invisible guardian standing between innovation and vulnerability."

Artificial intelligence is changing cybersecurity on both sides. Organizations are using AI to identify suspicious activity, analyze large amounts of data, and respond to threats more quickly. Cybercriminals, however, can use the same technology to create convincing phishing messages, impersonate trusted individuals, and automate parts of their attacks.

The concern is not simply that attackers have gained access to more powerful tools. AI can reduce the effort required to personalize attacks and make fraudulent communications appear more credible. Research from IBM, CrowdStrike, the World Economic Forum, and other organizations illustrates how these developments are changing cybersecurity. It also reveals a second problem: many businesses are adopting AI faster than they are establishing the controls needed to use it safely.

How Much AI Is Actually Appearing in Cyberattacks?

IBM's Cost of a Data Breach Report 2025, based on breaches experienced by 600 organizations worldwide between March 2024 and February 2025, found that 16% of the breaches studied involved attackers using AI. Among these AI-assisted attacks, 37% involved AI-generated phishing or other communications, while 35% involved deepfake impersonation.

The findings became more concerning in IBM's 2026 report. Published in July 2026, the study reported that one in four malicious breaches involved AI, representing a 56% increase over the previous year. IBM also reported an average cost of approximately $6 million for AI-enabled breaches, compared with $4.99 million for breaches overall.

These figures suggest that AI is becoming a more important part of the cyberthreat landscape. However, they should be interpreted within the limits of IBM's research, which examines a defined sample of organizations that experienced data breaches rather than every cyberattack worldwide.

There is also an important distinction between attacks conducted with AI and security incidents involving an organization's own AI systems. A criminal might use AI to impersonate an executive, for example, while a separate incident might involve an insecure AI application exposing confidential information. Both are cybersecurity concerns, but they require different preventive measures.

The growing use of unauthorized AI tools creates another risk. Known as shadow AI, this occurs when employees use AI services without the organization's approval or appropriate oversight. Employees might enter confidential business information into an external tool without understanding how that information is stored or processed.

IBM's 2025 research found that one in five organizations reported a breach involving shadow AI. Organizations with high levels of shadow AI experienced average breach costs $670,000 higher than organizations with low or no shadow AI exposure. The same report found that 97% of organizations reporting breaches involving their AI models or applications lacked proper AI access controls.

Together, these findings highlight two responsibilities: organizations must defend against criminals using AI, and they must ensure that their own adoption of AI does not introduce avoidable security weaknesses.

Why AI-Generated Phishing Can Be More Convincing

Phishing works by manipulating trust. Attackers impersonate banks, colleagues, suppliers, or managers to persuade people to reveal confidential information, open malicious links, or authorize payments. Traditional phishing messages may contain spelling mistakes, generic greetings, or suspicious requests, but well-crafted attacks can be difficult to recognize even without AI.

Generative AI makes it easier to produce personalized messages, adapt language to different audiences, and use publicly available information to make communications appear relevant. It can also help attackers create messages in languages they do not speak fluently. These capabilities lower some of the barriers to producing targeted attacks at scale.

A 2026 study published in Expert Systems with Applications examined AI-generated spear-phishing messages with 101 participants. The control group, which received generic scam messages, recorded a 12% click-through rate. Fully automated AI-generated phishing messages achieved 54%, while AI messages created with human assistance achieved 56%. Human experts also achieved a 54% rate in the experiment.

The results show that AI-generated messages can be highly persuasive under certain conditions. Nevertheless, the study involved a relatively small group of participants and a controlled experimental setting. Its findings should not be presented as the average success rate of all phishing campaigns.

Voice-based social engineering presents a related challenge. In these attacks, criminals may impersonate an employee, senior executive, or technology support worker to obtain credentials or persuade someone to transfer money. CrowdStrike's 2025 Global Threat Report recorded a 442% increase in observed voice-phishing operations between the first and second halves of 2024.

This figure describes a change in CrowdStrike's observations during a specific period. It does not establish that every form of voice fraud increased at the same rate or that AI alone caused the increase. Even with that qualification, it highlights the importance of protecting employees against impersonation and fraudulent requests.

Deepfakes and the Limits of Human Judgment

Deepfakes use AI to generate or manipulate images, videos, and audio so that a person appears to say or do something that never happened. In cybersecurity, these tools can be used to impersonate executives, imitate customers, or create false evidence to support fraudulent requests.

One widely reported example occurred in Hong Kong in 2024. An employee at engineering company Arup was deceived during a video conference in which the other apparent participants, including a person impersonating the company's chief financial officer, were deepfakes. The employee subsequently transferred approximately US$25.6 million to fraudulent accounts. The incident demonstrated how an attack could exploit the authority of senior executives and the apparent credibility of a live video meeting.

Research from identity-verification company iProov provides further evidence of the difficulty people can have identifying manipulated content. In a February 2025 study involving 2,000 consumers in the United Kingdom and United States, only 0.1% correctly distinguished all the real and fake images and videos presented in the test.

This statistic requires careful interpretation. It does not mean that only 0.1% of people can identify any deepfake in every situation. It means that only 0.1% of participants correctly identified all the test stimuli. Nevertheless, the result illustrates why visual or auditory judgment alone is an unreliable foundation for high-risk decisions.

The solution is not to assume that every video call or voice message is fake. Instead, organizations should change how they verify sensitive requests. An employee asked to transfer money should confirm the request through an independently established contact method. Changes to supplier bank details should undergo separate verification, and high-value payments should require documented approval from authorized personnel.

These procedures remain useful even when a deepfake is convincing. They do not depend on an employee being able to recognize every manipulated image or voice.

Governance Is Lagging Behind AI Adoption

Cybersecurity depends on more than purchasing advanced software. Organizations also need clear rules about how technology can be used, who can access sensitive information, and how suspicious activity should be handled.

IBM's 2025 report found that 63% of breached organizations either lacked an AI governance policy or were still developing one. Among organizations that had experienced breaches involving their AI models or applications, 97% lacked proper AI access controls (IBM, 2025). These findings suggest that some organizations have introduced AI without establishing sufficient oversight.

The World Economic Forum's Global Cybersecurity Outlook 2025 identified a similar gap between adoption and preparedness. Although 66% of surveyed organizations expected AI to have the most significant impact on cybersecurity in the coming year, only 37% reported having processes to assess the security of AI tools before deployment.

These findings do not prove that every organization without a formal policy will experience a breach. They do, however, indicate that AI security requires greater attention during procurement, implementation, and day-to-day use.

A practical governance framework should specify which AI tools employees may use, what information they may share, and who is responsible for approving new applications. Organizations should also review access permissions, monitor for unauthorized tools, and establish procedures for reporting possible data exposure.

AI governance should not become an obstacle to every new technology. A complete ban on useful tools may encourage employees to use them secretly. A better approach is to provide approved alternatives, explain the risks clearly, and set boundaries that protect sensitive information while allowing responsible innovation.

Why Small Businesses May Face Greater Consequences

The impact of a cyberattack depends not only on its financial cost but also on the resources available to the affected organization. Large companies may have dedicated security teams, incident-response arrangements, legal support, and sufficient reserves to manage disruption. Smaller businesses often operate with fewer resources and rely on a limited number of employees or digital systems.

The World Economic Forum's 2025 report found that 35% of small organizations believed their cyber resilience was inadequate, a proportion seven times higher than in 2022. By comparison, the share of large organizations reporting insufficient cyber resilience had nearly halved.

This difference matters because a successful attack can interrupt sales, delay payments, expose customer information, and damage relationships with suppliers. A small business that relies on a single accounting platform or online sales system may struggle to operate while that system is unavailable.

It would be misleading to claim that every small business faces the same likelihood of an attack or that a single loss amount will inevitably force a company to close. The consequences depend on its financial position, the information exposed, the duration of the disruption, and the effectiveness of its recovery arrangements.

However, smaller firms can take practical steps without building an expensive security department. Multi-factor authentication, regular software updates, secure backups, limited user permissions, employee training, and a documented recovery plan can help reduce exposure. Businesses should also know whom to contact if an incident occurs and how they will continue essential operations while systems are restored.

Cybersecurity guidance should recognize that organizations have different resources. Affordable, manageable controls can be especially valuable to smaller firms that cannot absorb prolonged disruption.

What Organizations Should Do in Practice

The evidence points to several priorities that organizations can put into action.

Strengthen identity and access controls. Multi-factor authentication, least-privilege access, and regular permission reviews can make it harder for attackers to exploit stolen credentials or misuse compromised accounts. AI applications should also receive appropriate access restrictions rather than automatically inheriting broad permissions.

Verify sensitive requests independently. Payment instructions, changes to supplier bank details, password resets, and requests for confidential information should be confirmed through trusted channels. A familiar voice, realistic video, or professional-looking email should not be sufficient evidence for a high-risk action.

Train employees using realistic examples. Awareness programmes should explain AI-assisted phishing, impersonation, and suspicious requests. Employees should be encouraged to report unusual communications rather than being expected to identify every deepfake correctly.

Govern the organization's own AI use. Businesses should maintain an approved list of tools, define rules for handling confidential information, and monitor for unauthorized applications. Third-party AI services should be assessed for relevant security and privacy risks before sensitive data is shared with them.

Use AI defensively with appropriate oversight. AI-powered security tools can help detect unusual activity, prioritize alerts, and support investigations. However, they need testing, human oversight, and clear procedures for responding to their findings. Organizations should measure whether these tools improve detection and response rather than assuming that AI automatically makes a system secure.

Prepare for recovery. Secure backups, tested incident-response plans, clear reporting channels, and defined responsibilities can limit the consequences of a successful attack. Prevention matters, but organizations also need to know how they will recover when preventive measures fail.

Cybersecurity Must Depend Less on Trust Alone

The most important lesson from these findings is that cybersecurity is not simply a contest between increasingly powerful machines. It is also a question of whether organizations have designed procedures that remain reliable when people can be deceived by convincing digital impersonations.

The iProov study highlights the limitations of human judgment in a specific deepfake-identification test. The phishing experiment shows that personalized AI-generated messages can achieve high click-through rates under controlled conditions. Neither result should be treated as a universal measure of human vulnerability, but both support the need for stronger verification practices.

Governance deserves equal attention. When businesses introduce AI without appropriate access controls and oversight, they can create security risks within their own operations. Some attacks require advanced defensive technology, but other weaknesses can be addressed through established measures such as restricted permissions, employee training, independent approval, and secure handling of data.

Small businesses should not be left out of this discussion. A security strategy designed for a large corporation may be too expensive or complicated for a small firm. Practical guidance should therefore emphasize affordable controls, accessible support, and recovery arrangements that match the organization's resources.

AI will continue to create opportunities for both attackers and defenders. Organizations cannot rely on employees to recognize every fake, nor should they assume that security software will prevent every breach. A more dependable approach combines appropriate technology with clear accountability, independent verification, and preparation for recovery.

In the age of AI, the objective is not to eliminate trust. It is to ensure that important decisions are supported by evidence and reliable processes rather than appearances alone.

Sources

CrowdStrike. (2025). 2025 Global threat report. https://www.crowdstrike.com/en-us/blog/crowdstrike-2025-global-threat-report-findings/

Heiding, F., Lermen, S., Kao, A., Mayrink Verdun, C., Schneier, B., & Vishwanath, A. (2026). Evaluating large language models’ ability to automate spear phishing. Expert Systems with Applications, 314, Article 131546. https://doi.org/10.1016/j.eswa.2026.131546

IBM. (2025, July 30). IBM report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls. https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications-97-of-which-reported-lacking-proper-ai-access-controls

IBM. (2026, July 29). IBM study: One in four malicious breaches are AI-enabled, costing companies $6 million on average. https://newsroom.ibm.com/2026-07-29-ibm-study-one-in-four-malicious-breaches-are-ai-enabled-costing-companies-6-million-on-average

iProov. (2025, February 12). iProov study reveals deepfake blindspot: Only 0.1% of people can accurately detect AI-generated deepfakes. https://www.iproov.com/press/study-reveals-deepfake-blindspot-detect-ai-generated-content

World Economic Forum. (2025, January 13). Global cybersecurity outlook 2025. https://www.weforum.org/publications/global-cybersecurity-outlook-2025/