Some of the strongest warnings about artificial intelligence are no longer coming from critics standing outside the industry. They are coming from the people building the technology itself.
Anthropic chief executive Dario Amodei has warned that the development of frontier AI may need to slow, arguing that safety systems and independent oversight need time to keep pace.
Sam Altman, CEO of OpenAI; Demis Hassabis, CEO of DeepMind; and Elon Musk have also backed stronger safeguards around increasingly capable systems, while current and former AI researchers have issued progressively sharper warnings about what could happen as models become more autonomous.
Stay up to date with the latest news. Follow KT on WhatsApp Channels.
The language can sound extreme. Some researchers are discussing systems that might one day improve themselves, operate with less human supervision or become difficult to control.
Others are concerned about risks already visible today, from AI-assisted cybercrime to agents that can write code, use tools and perform multi-step tasks. For the public, those concerns are easily collapsed into one alarming phrase: AI risk.
Some dangers already affect residents and businesses. Others are emerging at the frontier of AI development. The most severe scenarios remain hypothetical. That distinction matters in the UAE, where artificial intelligence is being adopted rapidly across government, businesses and everyday life.
Some of the risk is already here
Cybersecurity offers one of the clearest examples of what AI risk looks like today.
Yassin Watlal, CrowdStrike’s Senior Director of Systems Engineering for the Middle East, Turkey and Africa, told Khaleej Times that AI had already made some forms of cybercrime easier to carry out.
“AI has dropped the barrier of entry for a lot of the threat actors,” he said. “The less sophisticated actors now can do things that they were not able to either do, or it would take a lot of time.”
That does not necessarily mean criminals are inventing entirely new attacks. In many cases, AI is making old ones faster and more convincing.
Phishing messages can be written in better language and adapted to the person or country being targeted. Voice can be used for impersonation. Malware can be produced more quickly. The steps after an attacker gains access can also happen much faster.
At the same time, attackers increasingly do not need to force their way into a system if they can steal the identity of somebody who already has permission to enter. “They will want to log in, not break in,” Watlal said.
CrowdStrike says the average time taken by cybercriminals to move from an initial compromise into other parts of an organisation fell to 29 minutes in 2025. Its fastest observed case took just 27 seconds.
For a security team, that compression matters. An attack can begin moving through a network before a person has finished investigating the first suspicious alert.
Watlal said AI is therefore being used on both sides. Attackers can automate parts of their operations, but defenders increasingly need AI to investigate alerts and respond at similar speed.
That is one reason experts caution against treating AI safety solely as a future problem involving hypothetical superintelligence. Some effects are already measurable.
The bigger change is that AI is beginning to act
The next level of concern is more difficult to see. Most people first experienced generative AI through a chatbot. A person typed something, the model produced an answer, and the interaction ended there.
The industry is now increasingly focused on agents.
Instead of simply responding, an AI agent can be given an objective and allowed to take several steps towards completing it. Depending on the system, that may involve writing code, searching through information, calling other software, analysing files or performing actions using external tools.
That creates a different safety problem.
A chatbot can give a wrong answer; it ends by that. An agent can potentially turn a wrong decision into an action.
That ambition does not mean AI will suddenly be making every government decision independently. But it does make questions around permissions, human oversight and system security much less theoretical.
What can the agent access? What actions can it take without asking a person? What happens if it misunderstands a task? How quickly can a human stop it?
Those are increasingly practical questions for companies as well. Watlal said businesses need to think not only about defending themselves against criminals using AI but also about securing the AI systems they deploy internally.
“What can we ask the AI to do safely and then put guardrails around it?” he said. That becomes more important as AI systems are connected to company data, internal tools and automated workflows.
Then comes the fear that has not happened
Some frontier AI researchers are concerned about a future in which systems become capable enough to contribute substantially to developing their own successors.
AI already helps engineers write code, analyse research and perform parts of the work involved in building newer models. The concern is about what could happen if that assistance became significantly more autonomous and systems began accelerating the development of even more capable systems.
That is where discussions about recursive self-improvement, superintelligence and loss of human control begin. There is no scientific consensus that such a process will occur, when it might happen or whether future AI systems would behave in the catastrophic ways some researchers fear.
And it is exactly that uncertainty that is precisely dividing the industry.
One side argues that the potential consequences could be so serious that waiting for proof would mean waiting too long. The other warns against slowing a technology with enormous economic and scientific potential because of scenarios that remain speculative.
Even among people who agree that stronger safeguards are needed, there is disagreement over how far to go.
Calls to “slow down” AI development generally do not mean turning existing systems off or asking people to stop using them.
The proposals being discussed include stronger independent evaluations, more time between major capability jumps, better testing before release, greater coordination between frontier laboratories and clearer rules around what highly capable systems should be allowed to do.
What should UAE residents actually do?
For most residents, the immediate response is far less dramatic than the warnings about future AI might suggest. People do not need to stop using AI because researchers are debating long-term risks. But the current risks deserve attention precisely because they are less spectacular.
AI-assisted fraud, impersonation and cyberattacks are already happening. Watlal recommends enabling multi-factor authentication on personal accounts, checking unexpected requests through another channel and slowing down before approving a login or clicking a link.
“Things go so quickly in life that quite often we respond to something that then we regret, or we approve something that then we regret,” he said.
For businesses, the issue goes further. Companies adopting AI agents need to decide what information those systems can reach, what actions they can carry out and where a human must remain involved.
Watlal said the human role remains critical, particularly for judgement, context and accountability. The broader AI debate can therefore be separated into two questions.
The first is already here: how should people protect themselves from criminals and organisations using increasingly capable AI today? The second is much harder: how much independence should society eventually give the AI systems themselves?
People building the technology don’t agree on the answer, but what has changed is that some of them increasingly believe society should start asking the question before the most powerful systems arrive.
Source: Khaleej Times