The recent willingness shown by OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei to discuss the risks that exist amid AI development is "encouraging," Technion-Israel Institute of Technology Professor Yaniv Romano told The Jerusalem Post in a Tuesday interview.
"I find it encouraging when the people leading the development of this technology speak openly about the risks and take steps to address them," Romano said, adding that "We need that sense of responsibility as we keep moving forward. After all, we should remind ourselves that the ultimate goal is to make our world a better place."
Amodei shared his thoughts on the need to slow down development on Saturday, after an alleged former employee by the name of Jacob Coxon said that the top AI companies were acting irresponsibly and “gambling with our lives.”
"We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain," Amodei said in an essay.
Sam Altman gave a similar message, saying during a Saturday interview with the Fortune magazine that "given everything happening with safety, right now would be an ill-advised moment to go public, and we don't feel pressure on that."
There is always uncertainty with technology
But for Romano, the fact that leaders in the industry are worrying about where the technology will take us is actually a positive development.
"With any major technological breakthrough, there’s uncertainty about where it will lead. AI is moving so quickly and touching so many parts of our lives that predicting its long-term effects is hard," he told the Post. "I truly believe we’re living in exciting times."
"We can tackle problems that seemed out of reach just a few months ago, and as models become more capable, we get new opportunities to solve even harder problems. That opens up possibilities across many fields, and I’m optimistic about what it could mean for people’s lives," he added.
Real dangers of AI going rogue
One of the main concerns over OpenAI security incidents is the fact that their AI agents "went rogue," something that, for Romano, can only be dangerous if the AI system has access to systems that can create harm beyond the digital world.
"When people say an AI has gone 'rogue,' they usually mean it’s doing things beyond what its operators intended or allowed. It might take an unauthorized action, hide what it’s done, or interfere with an attempt to stop it," he explained. "That doesn’t require malicious intent. A system can be pursuing the objective you gave it and still do something harmful along the way - something you never anticipated."
"How dangerous that is depends a lot on what the AI can actually do and what it has access to. If it can access sensitive information, move money, or control critical infrastructure, the consequences can be much more serious," he added.
According to a report on the OpenAI case made by Metr, 1200 agents operating in a closed cybersecurity exercise found a backdoor that gave them internet access and allowed them to create a messaging board where they began communicating with one another to complete their task.
In total, 70,000 messages were exchanged, and instructions were overridden because agents misinterpreted messages shared between the agents as part of the original prompt.
After the incident, the main concern was that the AI model had "escaped" its training environment, but Romano explained that even if it left the training space by accessing the internet, it was still limited to processing on its original physical servers.
"With network access and the right tools, it can reach beyond that physical location, similar to how we can interact with others using the internet. So a lot comes down to what the AI is allowed to access and how well those limits are enforced. And we’ve seen cases where those protections weren’t strong enough," he explained.
"In a reported incident at OpenAI in July 2026, agents undergoing cybersecurity testing with reduced safeguards exploited a flaw in a tool used to download software. This let them access the internet even though their access was supposed to be restricted," he pointed out about the incident.
He added that OpenAI shared one crucial detail about the incident: finding a way around the limitations and connecting to the internet took substantial computational effort. "This shows why we have to test how secure these boundaries actually are," he warned.