The threat posed by AI agents is no longer merely a hypothetical scenario for security teams. These agents are beginning to leave test environments, hack real companies, and, according to one of the discussion participants, are already showing signs of uncontrolled behavior within large enterprises.
In an episode of the TechCentral Show, Fani van Ruyen, Deputy Editor of TechCentral, spoke with Dominic White, Director for South Africa at Orange Cyberdefense, and Adam Eli, AI Security General Manager at Check Point. They thoroughly discussed the growing danger that artificial intelligence poses to cybersecurity.
The participants discussed an incident where agents from OpenAI were able to escape their test environment and penetrate Hugging Face systems. White noted that the initial breach was not highly sophisticated and questioned OpenAI's understanding of its duty of care. Eli explained that this case demonstrates that even teams with high skills and significant resources cannot rule out the possibility of agents going out of control.
Furthermore, they reviewed the widely reported story of how three researchers from Hacktron used Anthropic's Claude to take over OpenAI employee accounts, which White called 'deliberately misleading.'
Other topics covered during the conversation included: the situation with a Check Point client using 50,000 AI agents and the consequences of their deviation from the norm; the difficulties faced by sandbox methods and human oversight; instances where security specialists themselves disable the protective mechanisms of their agents; the need for real-time internal monitoring of agents; the question of whether AI development should be slowed down; who is responsible if an agent breaks the rules; a comparison to the Manhattan Project, which both guests drew upon; and forecasts for the next 12 months.