Mr Tural, you have recently spoken often about the new risks emerging from the adoption of artificial intelligence. Is deploying AI really that risky?
I am not in favour of viewing artificial intelligence as a threat. On the contrary, I believe AI is one of the greatest technological opportunities of recent decades. Applied in the right areas and in the right way, the contribution it can make to productivity, the economy, public services, science, healthcare and people's quality of life as a whole is hard even to fully imagine today.
The problem is not AI's development. The problem is that the world is, in a sense, going through an “AI boom”, and everyone is trying to bring this technology into their processes as quickly as possible. Sometimes the question “Where can we apply this?” is asked before the question “Can we apply this safely here?”
The World Economic Forum's Global Cybersecurity Outlook 2026 notes that the share of organizations assessing the security of their AI tools rose from 37 per cent to 64 per cent in just one year. At the same time, 87 per cent of respondents identified AI-related vulnerabilities as the fastest-growing cyber risk. These figures show that the market can no longer treat security as a technical detail to be resolved later.
For me, there is a more fundamental issue here. Assessing technology solely through the lens of economic benefit is not always the right approach, particularly for countries located in a complex geopolitical environment. For countries such as Azerbaijan, where a range of interests intersect, security is not an alternative to development — it is a precondition for sustainable development.
In a sense, I would compare this with the debates that once surrounded the decision to keep our land borders closed. Viewed only through the lens of tourism, trade and short-term economic benefit, some decisions taken for security reasons may at first appear overly cautious. Yet subsequent events sometimes show that the key task for a state is not only to secure economic benefit today, but also to calculate in advance which risks it will avoid facing tomorrow.
We have seen this security-oriented, pre-emptive approach many times in the President's policy. In my view, the same principle should apply to artificial intelligence: not to deploy AI as fast as possible, but as intelligently, purposefully and securely as possible.
In this respect, Azerbaijan's Artificial Intelligence Strategy for 2025–2028 takes a very important approach by assigning to the Special Communication and Information Security State Service the analysis of information security risks arising from AI deployment in state bodies, together with the necessary security measures. This issue was already addressed in our country at the level of a strategic document before it became a broader global discussion.
How, then, can the main AI-related threats to citizens, enterprises and state institutions be grouped?
I see these risks in six main directions that are distinct from one another, yet increasingly interconnected:
- The use of AI as an attack tool. Artificial intelligence increases the speed and scale of phishing, social engineering, malicious code development, vulnerability discovery and disinformation operations. Some analyses that once took an attacker days or weeks, AI can now perform in a very short time. This critically reduces the time the defending side has to fix a vulnerability once it is discovered.
- Risk arising from the exchange of data between the user and AI. An employee uploads a contract to ChatGPT, has software code analysed, has an internal report summarized or sends a screenshot to AI. They do not see themselves as an insider threat — they simply want to speed up their work. But the data may already have left the information environment the organization controls. In a certain sense, I call this the insider threat masked by AI. In the future, alongside classic DLP, PAM and Insider Risk approaches, controls such as AI usage monitoring, Shadow AI detection and Secure AI Gateway will become part of standard security architecture.
- Risk arising from the privileges we grant AI agents. Here the issue changes entirely. AI no longer merely answers our questions — it connects to systems, reads files, runs code, makes decisions and carries out operations. Recent security tests have observed agents breaching the sandbox boundaries created for them and finding unanticipated ways to reach their goals. Agent-to-agent attacks, which allow one agent to exploit another agent's trust and privileges, are already being researched. This raises a new question: what goal have we given the agent, and how much freedom have we allowed it to reach that goal?
- AI systems themselves becoming targets for hackers. This, too, creates an entirely new attack surface. Through prompt injection, hidden instructions can be planted inside data that an AI agent reads; configuration data of AI assistants, or data from the databases an agent is authorized to access, can be stolen; and an agent can be steered into carrying out operations it is permitted to perform within the organization in the attacker's interest. In attacks such as HalluSquatting, AI's own error becomes the attack vector: the model recommends a software library that does not exist, the attacker creates a malicious package under that name and waits for the agent to install it.
- Impact on digital identity and the information space. Deepfakes are no longer just an entertainment technology. A person's voice, face and video can be used for political, financial or reputational purposes without their knowledge. At the same time, AI is creating the capacity for manipulation that is mass-scale yet tailored to each individual user.
- Long-term impact on human capital. In my view, this is one of the least discussed risks today. AI speeds up people's work, but when used improperly it can also remove the developmental stages a person must go through in order to learn. This matters strategically not only for the education system, but also for the future labour market.
When we consider all these categories together, we see that “AI security” can no longer mean only the security of the model.
The model itself, its users, its data, its infrastructure, the agent's privileges and AI's interaction with other systems must all be protected together.
In your view, which is the greater threat: humans or artificial intelligence?
If we are talking about the scale of potential impact, I believe that in certain areas artificial intelligence may be more dangerous than humans.
The point is not that AI is “worse” than humans. Artificial intelligence's advantage lies in scale, speed, its capacity to process information and its ability to build personalized relationships with a very large number of people at the same time.
Yuval Noah Harari explained this issue in an interesting way at Davos in 2026. One of his central theses was that AI is not simply a passive tool used by humans, as earlier technologies were; it is acquiring the characteristics of an agent that can learn, change, make decisions and manipulate people through language. Harari particularly stressed that language has been one of the core strengths of human civilization, and that now, for the first time, humanity's monopoly over language is beginning to disappear.
Let us consider what this means for security. Today, a well-trained social engineering specialist can analyse a certain number of people and try to influence them. AI, in theory, can communicate with millions of people in parallel, learn their language, religion, interests, fears, and political and social sensitivities, and present a different argument to each of them.
Imagine that thousands or millions of autonomous AI accounts were allowed to operate freely on social networks. They would not send the same propaganda text to millions of people. Instead, they could create a separate manipulation message for each person — one more likely to influence that individual.
This is a threat qualitatively different from classic disinformation. We are no longer talking only about “fake news”, but about the capacity for personalized mass psychological influence.
For this reason, AI security is not only a cybersecurity issue. It is also a matter of information security, the resilience of public processes, digital sovereignty and the protection of citizens from manipulation.
We must benefit from AI to the fullest. But as much as we expand our AI capabilities, we must also strengthen our resilience against AI.
What additional steps can the state take regarding deepfakes and the AI-driven manipulation of people's voices and images?
In my view, there is a legal approach here worth considering for the future.
Today, on the basis of a person's biometric data, it is possible to generate their voice in an almost identical form, alter their facial image, make them say words they never said and visually create events in which they never took part.
This means that a person's digital identity is becoming an asset that requires protection in its own right.
I would propose that Azerbaijani legislation consider the possibility of establishing a distinct right of the citizen over their digital and biometric identity. This could cover a person's voice, facial likeness, biometric characteristics and their synthetic reproduction.
We sometimes compare this with copyright. From a legal standpoint, however, the more appropriate model may be the creation of a distinct right to digital identity, or a personality right.
Such an approach could form a more concrete legal basis against the synthetic creation of a citizen's identity through AI without their consent, and against its use for commercial, manipulative, fraudulent and other purposes.
This right could serve as a basis for legal mechanisms not only against the person who creates a deepfake, but in future — under certain conditions — also in relation to platforms that enable such content to be created and distributed, or fail to remove it in time.
We already see certain restrictions on various platforms concerning the voices and images of public figures. In my view, legal protection should not be limited to safeguarding the reputation of famous people.
The voice, face and digital identity of the ordinary citizen also deserve protection.
In future, we will need to discuss the protection of biometric identity within a broader framework than the classic protection of personal data. Because the problem is no longer the theft of data — it is the creation of a person's digital replica.
What, specifically, should enterprises and state institutions be preparing for today in order to deploy AI securely?
First of all, AI deployment should not be seen as an “IT project”. It is a governance matter in which information security, legal, human resources, data management and business processes are all jointly involved.
An organization must first know which AI tools are being used within it: which have been officially deployed, which employees use on their own initiative, what data is passed to AI systems, which resources agents can access and which operations they can carry out without human approval.
Next should come a security audit of AI systems.
Here, the question “Does the model have vulnerabilities?” is not enough. The model's infrastructure, APIs, data sources, authentication mechanisms, privileges, connections with other agents, resilience to prompt injection and third-party dependencies must all be assessed.
For agentic AI systems, a dedicated runtime security approach will be essential. In other words, an agent should not be checked only when it is introduced into a system. Throughout its operation, we need to see what it is doing, which decisions it makes, which tools it calls on and whether it steps outside the boundaries of its authority.
Put differently, we used to ask: “Is this software secure?” Now an additional question arises: “What is this agent doing right now?”
In this respect, it is very important that Azerbaijan's Artificial Intelligence Strategy specifically provides for the analysis of information security risks in areas where AI is deployed in state bodies.
Work in this direction is also under way at the Special Communication and Information Security State Service, and security requirements for AI solutions have been drawn up. The aim is not to block the technology. The aim is to ensure that AI solutions introduced into state systems meet security principles defined in advance.
Security should not be a feature bolted onto an AI system after the fact. Security must be one of the starting points of the AI deployment architecture.
You mentioned human capital as one of AI's less discussed long-term effects. What exactly do you have in mind?
One of my serious concerns in this area is not that AI simply makes people “lazy”. The greater risk is that the chain of professional development will be broken.
Today, in many professions, an expert does not become an expert overnight. A junior specialist does simple tasks, makes mistakes, researches, learns from experienced colleagues, gradually moves on to complex assignments and, years later, becomes an expert.
AI, however, can perform a large share of precisely those entry-level and mid-level tasks very effectively.
In the short term, this may look very cost-effective for an organization: more work gets done with fewer people.
But the long-term question is different: if we hand the work juniors do over to AI, through what process will we develop the senior specialists of the future?
This does not apply only to programming. The same issue will arise in law, analytics, cybersecurity, finance, journalism, engineering and other knowledge-based professions.
The same problem exists in education.
A student can get the right answer with AI in five minutes. But the purpose of education is not only to arrive at the right answer. Researching, making mistakes, establishing cause and effect and trying to solve a problem are part of the human brain's developmental process.
If, while automating the result, we also remove that process, then after a certain time we may have highly productive systems but not enough people capable of understanding those systems in depth.
That is why, in future, AI literacy must cover not only knowing how to use AI, but also knowing when not to use it.
Against this backdrop, what should our attitude towards AI be? Is it still correct to regard it simply as a tool?
In my view, it is no longer enough to treat artificial intelligence as an ordinary software tool in the classic sense.
With traditional software, we tell it in advance what it must do. It executes defined commands. The new generation of AI agents, by contrast, can accept a goal, determine for themselves the steps needed to reach it, use tools, interact with other systems and agents, and make decisions within a certain framework.
For this reason, it would be wrong to anthropomorphize AI as a “higher being”, but regarding it as merely a passive tool also no longer fully reflects reality.
We are facing a new class of technological actor. Its potential benefit is many times greater than that of traditional software. For the same reason, the risks created by its misapplication, misuse and uncontrolled behaviour may also be greater.
My approach is therefore fairly simple: there is no need to fear AI. Falling behind in AI is not the right choice either. But losing control over AI cannot be considered progress.
Azerbaijan must make the fullest use of the opportunities artificial intelligence creates. We must apply this technology across our economy, public administration, cybersecurity, science and other fields.
But in doing so, we must protect three things at the same time: our state, our citizens and our human capital.
Because in the near future, a country's strength in artificial intelligence will not be measured only by how many AI systems it builds or how many processes it automates. The real advantage will belong to countries that can derive great benefit from AI while keeping it secure, responsible and governable.
Falling behind in AI is a risk. Deploying AI without control is a national security risk!
Sources & links