Three security experts from different companies are reflecting on the risks of using autonomous AI agents
Michaela Stonová (OKsystem), David Doležal (Productboard), and Luboš Řádek (České Radiokomunikace) agree that, in the wrong hands, agents can become a dangerous weapon.
Artificial intelligence in companies is evolving from chat assistants to autonomous agents—and with that comes a new type of security risk. While traditional IT relies on antivirus software and firewalls, no comparable off-the-shelf tools currently exist for agent-based systems. Protection therefore does not rely on a single technical solution, but on prevention, clearly defined processes, and the gradual establishment of barriers. The real problem today is no longer a sci-fi scenario, but entirely mundane issues: the reckless linking of work and personal accounts, or an agent with access to a calendar, email, and online banking.
“At OKsystem, we treat agents much like new employees—that is, we maintain records for them and grant them permissions appropriate to their work,” explains Michaela Stonová, noting how we work with multi-agent systems. “Irreversible or sensitive operations are always approved by a human; all steps and intermediate steps are continuously logged, and the security framework includes the ability to immediately stop an agent. We never deploy multi-agent systems into the production environment—given that our solutions are also deployed in critical infrastructure, we cannot afford to make any mistakes. A prompt alone is not enough to keep the agent within bounds; the barriers must be such that the agent cannot ‘physically’ overcome them.”
The choice of model is just as crucial. It’s not just about price and capabilities, but also about who operates the model, where the data is processed, and what security guarantees are provided. When it comes to personal data, processing within the European Union is key; we process truly sensitive data in locally operated models, and that data never leaves the company’s perimeter.
“AI agents aren’t just the next wave of automation. They are new virtual employees with their own ability to act—and to make mistakes. Organizations that deploy them without proper oversight are creating a security debt that could come back to haunt them when they least expect it,” Stonová concludes.
You can read the full article in Czech on the E15.cz website.