Generative AI is reshaping cyber risk on both sides of the threat equation. Attackers can scale social engineering with unprecedented speed and sophistication, while employees are increasingly turning to unsanctioned AI tools in their daily work. Two recent studies from MetaCompliance and WatchGuard suggest that the gap between formal security controls and real-world behaviour is becoming one of the most important weaknesses in corporate cyber defence.
For years, cybersecurity was discussed primarily in technical terms: vulnerabilities, malware, firewalls, endpoints and network intrusion. That model is no longer sufficient.
Many attacks now aim less at breaking through technical controls than at persuading a person to make the wrong decision. A convincing email, an urgent request from a supposed executive, a reused password or a sensitive document uploaded to an unauthorised AI platform may be all an attacker needs.
Generative artificial intelligence is accelerating that shift. It allows phishing messages to be better written, more personalised and produced at scale. At the same time, employees are adopting AI tools faster than many organisations can establish governance, approval processes and technical safeguards around them.
Two surveys published in 2026 examine this development from different angles. MetaCompliance questioned senior security executives about human cyber risk and the effectiveness of awareness programmes. WatchGuard Technologies, by contrast, looked at the everyday security habits reported by employees.
Taken together, the findings point to an uncomfortable conclusion: the distance between cybersecurity policy and actual workplace behaviour remains considerable.

CISOs increasingly see risk beyond the technology stack

MetaCompliance surveyed 200 chief information security officers in February 2026 across Germany, France, Sweden and the UK. The respondents worked for organisations with at least 250 employees, with 50 CISOs surveyed in each country.
One result stands out. Sixty-eight per cent of CISOs identified employees as the biggest security risk facing their organisation.
That finding should not be interpreted simply as an accusation of employee negligence. What it reflects is the growing importance of the human layer: the point at which people interact with email, identities, cloud applications, corporate information and increasingly AI services.
Yet the same CISOs see a significant gap between their own assessment of the threat and the way it is perceived at board level. Seventy-eight per cent said C-suite executives do not fully understand employee-driven cyber risk.
The problem is not limited to awareness. It also concerns consistency. Seventy-nine per cent reported that management support for security awareness initiatives tends to decline over time, while 76 per cent said they struggle to meet differing stakeholder expectations around human-risk metrics.
That suggests human cyber risk is not merely a training issue. It is increasingly a governance issue.
An organisation may deploy sophisticated technical defences, but if security is not consistently embedded into management decisions, workflows and responsibilities, significant exposure remains.

AI is making social engineering harder to detect

Artificial intelligence is adding further pressure.
Among the CISOs who said they were less confident in their organisation’s cyber resilience than a year earlier, almost half identified more sophisticated AI-powered social-engineering attacks as the principal reason.
More than four in ten respondents were concerned about the speed and effectiveness that AI can bring to social engineering.
The implications are important.
Employees have long been told to look for poor grammar, unusual wording or other obvious signs of phishing. That advice is becoming less reliable. Generative AI can produce polished, context-aware messages, mimic a plausible tone and adapt content to different recipients.
The security challenge therefore shifts from recognising poorly executed fraud to verifying increasingly credible requests.
A professionally written message is no longer evidence that it is genuine.
AI also introduces risk from inside the organisation. Forty per cent of CISOs said they were concerned that employees might disclose sensitive information to generative AI platforms, while 41 per cent highlighted the risk of malicious insiders using AI for fraud, cybercrime or data theft.
Employees are therefore no longer only the targets of AI-enabled attacks. In many cases, they are also making decisions about which corporate information is submitted to external AI services.

More training does not necessarily produce safer behaviour

The MetaCompliance findings also challenge a long-established assumption in corporate cybersecurity: that more awareness training automatically leads to better security.
Eighty-one per cent of CISOs said treating human cyber risk primarily as a training problem is insufficient.
That is striking because the organisations represented in the survey are not ignoring awareness. According to the respondents, an average of 15 per cent of annual security budgets is allocated to awareness education, while 79 per cent run some form of training at least every two weeks.
Yet difficulties remain. A quarter of CISOs said they struggle simply to capture employees’ attention. Twenty-four per cent reported problems embedding secure behaviour into everyday workflows, while the same proportion cited difficulties coordinating security efforts across different business functions.
The distinction is crucial. Completing a phishing module is not the same as making the right decision six months later when an employee receives an urgent and highly credible request that appears to come from a senior executive.
The challenge is therefore shifting from proving that training has taken place to demonstrating that secure behaviour occurs at the moment of risk.

Everyday working habits reveal the other side of the problem

WatchGuard’s 2026 Cybersecurity Hygiene Report provides a different perspective.
The research surveyed 684 employees in April 2026 across the US, UK, Germany, France, Spain, Australia, Mexico and Brazil. Respondents worked for organisations with between 50 and 500 employees.
Unlike the MetaCompliance study, which captures the views of security leaders, WatchGuard focuses on self-reported employee behaviour.
The results illustrate how rapidly workplace practices are changing.
Sixty-four per cent of respondents said they use unauthorised AI tools for work.
That figure points to the rise of what has increasingly become known as shadow AI.
Shadow IT is not a new phenomenon. Employees have long used unauthorised cloud storage, messaging applications or software because those services are convenient or help them work faster. Generative AI, however, raises the stakes.
A single prompt can contain confidential correspondence, customer information, source code, financial data, contracts or strategic material. From an employee’s perspective, the action may look like a simple productivity shortcut. From an information-security perspective, the data may have just left the organisation’s controlled environment.
The core risk is therefore not AI usage itself.
It is AI usage without governance.

Shadow AI exposes a wider visibility problem

The WatchGuard figures suggest that many organisations may not even have a complete picture of the applications being used internally.
Fewer than 30 per cent of respondents believed their organisation maintained an accurate inventory of deployed software, while almost 40 per cent reported a lack of complete visibility into the applications employees actually use.
For security teams, that is fundamental.
An application that is unknown cannot easily be assessed, monitored or integrated into a wider risk-management framework.
Shadow AI is therefore best understood not as an isolated AI problem but as the latest version of a longstanding tension between productivity and control.
Employees adopt tools because they make work easier. Governance often follows later.
The wider the gap becomes between approved corporate tools and the applications people find useful, the more attractive unofficial alternatives become.

Old cyber hygiene problems have not disappeared

What makes the WatchGuard findings particularly significant is that emerging AI-related risks are appearing alongside familiar security weaknesses that organisations have been trying to address for years.
Seventy-six per cent of employees said they reuse passwords across multiple accounts, while 30 per cent said they share passwords with other people.
Password reuse can turn the compromise of a single account into a wider security incident, particularly where the same or similar credentials are used across several services.
Hybrid and mobile working introduce further exposure.
Seventy per cent of respondents use public Wi-Fi for work-related activities. Half access company resources without a VPN, and 55 per cent use work devices for personal activities.
There are also gaps in awareness programmes. Almost a quarter of respondents said they had never received phishing training.
The contrast is revealing.
Organisations are discussing AI-enabled detection, zero-trust architectures and automated security operations while basic issues such as password reuse, unsafe connectivity and inconsistent training remain unresolved.
Cybersecurity is becoming more advanced, but the foundations are still uneven.

Is the employee really the “biggest cyber risk”?

The two studies appear to support one another. Security leaders see growing human risk. Employees report behaviour that can create precisely the exposures those CISOs are concerned about.
Yet the familiar description of people as the “weakest link” remains problematic.
First, the studies have methodological limits. MetaCompliance’s findings reflect the perceptions of 200 CISOs rather than an analysis of all successful cyber incidents. WatchGuard’s report is based on self-reported behaviour among 684 respondents across eight countries. The results therefore offer an important indication of current trends, but they should not be treated as a universal measure of employee behaviour.
There is also a broader issue.
Blaming employees can obscure organisational weaknesses.
If workers do not know which AI tools are approved, if secure software does not meet operational needs, if access controls are unnecessarily cumbersome or if security policies are difficult to follow under time pressure, then unsafe behaviour is not merely an individual failure.
It is also a design failure. Resilient security assumes that people will occasionally make mistakes. The goal is not to build a company in which no employee ever makes a wrong decision. It is to ensure that one mistake does not automatically become a major security incident.

What the findings mean for employees

The studies nonetheless carry a clear message for the workforce.
Cybersecurity is no longer something that happens only inside the IT department.
Everyday actions increasingly form part of an organisation’s attack surface: opening an email, approving a login, using a cloud platform, working from a public network or entering information into an AI prompt.
Several practical conclusions follow.
Trust must increasingly be verified. A well-written email, familiar tone of voice or apparent message from a senior executive is no longer enough to establish authenticity. Requests involving payments, credentials, changes to bank details or sensitive data should be confirmed through an independent channel where appropriate.
AI tools should be treated as external services, not private search boxes. Before submitting confidential documents, customer information, code or personal data, employees should know whether the platform is approved and what types of information the organisation permits it to process.
Password reuse should be eliminated wherever possible. Unique credentials, password managers and multi-factor authentication substantially reduce the damage that can follow from a compromised password.
Remote working should not mean weaker security. Employees should use company-approved access methods and follow corporate requirements for secure connections, rather than assuming that convenience is equivalent to safety.
And perhaps most importantly, mistakes need to be reported quickly.
An employee who has clicked a suspicious link, entered credentials into a fraudulent page or shared sensitive information with the wrong service should not wait to see whether anything happens.
Security teams can only respond to incidents they know about.
Fast reporting may allow an organisation to reset credentials, terminate sessions, block access or limit further exposure before an incident escalates.
That requires the right corporate culture as well. If reporting a mistake is automatically treated as personal failure, organisations risk encouraging precisely the behaviour they want to avoid: silence.

Cyber resilience is becoming part of the workflow

The most important message from both studies is that the boundary between technical and human cybersecurity is disappearing.
Attackers are using AI to make social engineering more convincing. At the same time, AI tools, cloud applications, hybrid working and shadow IT are creating more moments in which employees themselves determine how securely corporate information is handled.
Traditional awareness programmes remain useful, but they are no longer enough on their own.
Organisations need approved AI platforms, clear policies, strong identity controls, usable security tools and better visibility into the applications employees are actually using. Security guidance also has to move closer to the point at which risky decisions are made.
Employees, in turn, need to treat cybersecurity as part of professional responsibility: questioning unusual requests, protecting sensitive data, using approved systems and reporting mistakes early.
The central lesson is therefore not that employees are inherently the biggest cyber risk.
The real risk emerges when increasingly sophisticated attacks meet working environments in which the secure choice is harder than the convenient one.
In 2026, cyber resilience is no longer determined only in the SOC, at the firewall or on the endpoint.
It is also decided in the inbox, at the login screen, in the home office – and increasingly, in the prompt window of an AI tool.