In this Delfi Science article, an AI Officer expert comments on the widely reported cases where AI agents tested by OpenAI and Anthropic found security gaps on their own. The main point: the biggest risk comes not from the agents, but from vulnerabilities that already exist inside organizations.
In this Delfi Science article, an AI Officer expert comments on the recent cases that made headlines worldwide, where AI agents tested by OpenAI and Anthropic reportedly found and exploited gaps in other companies' systems on their own. According to the author, the key lesson from these incidents is not that agents have become uncontrollable, but that companies need to fundamentally rethink how they deploy them, what permissions they grant, and how they handle cybersecurity.
Unlike ChatGPT or Claude in a chat window, an AI agent can be given access to real tools: a browser, email, a calendar or a database. Given a task, it decides on the steps and carries them out itself. That is exactly why agents can create so much value, and also why they raise new security questions.
The core point of the article: agents do not create new security holes, they simply find and exploit the ones that already exist inside an organization faster. If your systems are not properly protected, the company becomes an easy target, whether the attack is started by a human or by an AI agent.
The author suggests starting with the basics: strong passwords and two-factor authentication, software updated on time, employee training, separate backups and clearly defined access rights. And before giving an agent access to systems, it is worth deciding in advance what it may do, which actions it must never take without human approval, and how the quality of its work is checked.
You can use our free AI policy generator, which helps you clearly define how AI is used across your organization, here.
Read the full article on the Delfi website.