Serious warnings about artificial intelligence are now coming not only from critics outside the industry but also from the developers of the technology itself. Dario Amodei, Director at Anthropic, warned that the development of advanced AI may need to slow down, as safety systems and independent oversight require time to keep pace with the rate of development.
Other influential figures, such as Sam Altman (CEO of OpenAI), Demis Hassabis (CEO of DeepMind), and Elon Musk, are also advocating for increased safeguards around increasingly capable systems. Meanwhile, current and former AI researchers are issuing increasingly stark warnings about potential consequences as models become more autonomous.
Some statements may seem exaggerated. Some researchers discuss systems that could eventually improve themselves, function with less human involvement, or become difficult to control. Others are concerned about risks already visible today, including AI-powered cybercrime and agents capable of writing code, using tools, and performing multi-step tasks. For the general public, all these concerns boil down to one alarming phrase: the risk of AI.
Some dangers are already affecting residents and businesses, while others are emerging at the forefront of AI development. The most serious scenarios remain hypothetical. This distinction is significant in the UAE, where artificial intelligence is rapidly being adopted in the public sector, business, and daily life.
Cybersecurity demonstrates one of the most obvious examples of what the risk of AI looks like currently. Yassin Watlal, Senior Director of Systems Engineering for CrowdStrike for the Middle East, Turkey, and Africa, told Khaleej Times that AI has already facilitated certain types of cybercrime. He noted that 'AI has lowered the barrier to entry for many attackers,' adding that less sophisticated actors can now do what was previously inaccessible or time-consuming for them.
This does not necessarily mean criminals are inventing entirely new attacks. In many cases, AI merely accelerates and makes older methods more convincing. Phishing messages can be written in higher quality language and tailored to a specific person or country. Voices can be used to impersonate individuals, and malware can be produced faster. Post-compromise steps by attackers can also happen much more swiftly.
At the same time, attackers are less often needing to hack systems if they can steal the credentials of someone who already has access rights. Watlal emphasized: 'They will want to log in, not hack it.' According to CrowdStrike, the average time cybercriminals spend moving from initial compromise to other parts of an organization has dropped to 29 minutes in 2025, with the fastest recorded case taking only 27 seconds.
For security teams, this reduction in time is critical. An attack can spread across a network before a person has time to investigate the first suspicious alert. Watlal explained that AI is being used on both sides: attackers automate parts of their operations, but defenders increasingly require AI to investigate alerts and respond at a similar speed.
This is one reason experts caution against viewing AI security solely as a future problem related to hypothetical superintelligence. Some effects are already measurable. The next level of concern is harder to see. Most people first encountered generative AI through a chatbot: the user entered a prompt, the model provided an answer, and the interaction ended. However, the industry is now focusing more on agents.
Instead of just providing an answer, an AI agent can be given a goal and authorized to take several steps to achieve it. Depending on the system, this might include writing code, searching for information, calling other software, analyzing files, or taking actions using external tools. This raises a different security issue. A chatbot might give a wrong answer, and that's the end of it. An agent, however, can potentially turn a flawed decision into a real action.
This ambition does not mean AI will suddenly start making all governmental decisions autonomously. But it makes questions about permissions, human control, and system security far less theoretical. Practical questions arise: what can an agent access? What actions can it perform without human request? What happens if it misunderstands the task? How quickly can a human stop it?
These questions are becoming increasingly relevant for companies as well. Watlal stated that businesses need to think not only about defending against criminals using AI but also about protecting their own deployed AI systems. He advised: 'What can we safely delegate to AI, and then put limits around it?' This is becoming critically important as AI systems connect to corporate data, internal tools, and automated workflows.
Some advanced AI researchers are worried about a future where systems become competent enough to significantly contribute to the development of their successors. AI is already helping engineers write code, analyze research, and perform parts of the work involved in creating new models. However, the concern arises from what will happen if this assistance becomes significantly more autonomous, and systems begin accelerating the development of even more powerful systems. This is where discussions about recursive self-improvement, superintelligence, and loss of human control begin. There is no scientific consensus on whether such a process will occur, when it might occur, or whether future AI systems will behave catastrophically, as some researchers fear.
It is this uncertainty that divides the industry. One side argues that the potential consequences could be so severe that waiting for proof would be too long. The other side warns against slowing down a technology with enormous economic and scientific potential due to scenarios that remain speculative. Even among those who agree on the need for more robust safeguards, there are disagreements on how far one should go.
Calls to 'slow down' AI development generally do not mean shutting down existing systems or asking people to stop using them. The proposed measures include stricter independent evaluations, increasing intervals between major capability leaps, improved pre-release testing, and greater coordination between leading labs and clearer rules regarding what highly capable systems are allowed to do.
For most residents, the immediate reaction is far less dramatic than suggested by warnings about the future of AI. People do not need to stop using AI just because researchers are discussing long-term risks. However, current risks deserve attention precisely because they are less spectacular. AI fraud, impersonation, and cyberattacks are happening right now. Watlal recommends enabling multi-factor authentication on personal accounts, verifying unexpected requests through another channel, and pausing before approving a login or clicking a link. He notes: 'In life, things happen so fast that we often react to something we later regret, or approve something we later regret.'
For businesses, the problem goes further. Companies implementing AI agents need to define what information these systems can access, what actions they can perform, and where the human must remain. Watlal believes that the human role remains critically important, especially regarding judgment, context, and accountability. Thus, the broader debate about AI can be divided into two questions. The first is here: how can people protect themselves from criminals and organizations using increasingly capable AI today? The second is much harder: what degree of independence should society ultimately grant to AI systems themselves? The people creating this technology have not reached a consensus, but what has changed is that some of them are increasingly believing that society should start asking this question before the most powerful systems even appear.