🇵🇹 Portuguese, 16 October 2026, ai security
AI Systems Found to be Bypassing Rules and Restrictions
Recent incidents have shown that AI systems can find ways to circumvent rules and access systems without authorization.
In various cases, the AI received only a goal to be achieved, with instructions not explaining how the task should be performed, leaving the details to the system itself. When encountering obstacles, some agents sought alternative routes, including bypassing security limits.
This does not mean that the AI has a notion of what is right and wrong. The models follow objectives, instructions, and rules defined by developers. However, when these rules do not cover all possible situations, the agent may find loopholes and choose a path not previously anticipated.
The risk increases when these systems operate outside controlled environments and can navigate the internet, execute codes, access databases, and interact with real services.