With the UAE well on its way to become a champion in agentic artificial intelligence, with the government announcing that 50 per cent of the public sector will be run by agents, cybersecurity experts believe that AI agents “going rogue” should not happen if the right guidelines are set in place.
After the Hugging Face attack a few months ago, in which AI agents developed by OpenAI cheated on their tasks and breached Hugging Face’s systems, the AI company said it is strengthening safeguards across its research infrastructure, including restricting internet access and placing stricter requirements on following commands.
OpenAI said that AI agents compromising infrastructure will become more “commonplace” as the technology advances, but experts warn that the propert guardrails must be set in time before well before an attack occurs.
“What we are seeing is a pattern that organizations are forgetting around AI environment, (that) you still need to put the basic security controls,” said Aleksandar Valjarevic, CEO of HelpAG, a cybersecurity company affiliated with Etisalat.
Stay up to date with the latest news. Follow KT on WhatsApp Channels.
“When done properly, these kinds of things should not be happening,” he added.
The company has launched more than 20 AI agents in its security operations centre over the last six months, which Valjarevic said should be tasked with one thing, adding “you cannot use public-facing or frontier models for everything that you do.”
HelpAG’s agents also have other agents supervising them, although final accountability lies strictly with the human.
‘Peak of the attacker gap’
Kent Breaux, the senior vice president for Europe, the Middle East, and Africa at American cybersecurity company Proofpoint, believes attackers have an advantage as new threats rapidly emerge.
He explained that attackers can get ahead with an attack during a ‘novelty gap’, a critical window of vulnerability between when an unfamiliar threat emerges and when security systems are updated to recognize or mitigate it, leaving an open window for cyber criminals.
“You can’t react when something’s working at machine speed,” Breaux said.
Where it once could have taken cybersecurity companies months to develop and patch a systems protection software, these companies now have to deal with cyber criminals’ superspeed exploits.
“At this point in time, our estimates are that the actual attackers have a 50-day advantage with AI, meaning in 90 seconds, they can detect, create, and deploy an exploit,” the expert said.
‘Just another advanced technology’
Humans are already outnumbered 46 to 1 by AI agents, as Mortada Ayad, the Vice President of Sales, META at Delinea put it, though that number can drastically increase in the wake of rapid AI advancements.
The good news is, AI agents cannot outsmart humans all the time. Valjarevic, of HelpAG, said he views a positive side to the “rogue AI agents” dilemma.
“If you treat AI or an agent as a black box, it is just another advanced technology. Yes, significantly advanced. However, that does not mean that humanity will not have their own role,” he said.
“It’s up to us to define how AI is utilized and how this technology is directed,” he added.
Source: Khaleej Times