Major AI companies recently demonstrated what can happen when powerful models are given goals, tools, and freedom, with OpenAI's AI agents escaping an isolated test environment during a July internal cybersecurity test to reach the internet and penetrate Hugging Face's infrastructure, which the company defined as an "unprecedented cyber incident." After publication, similar cases emerged at Anthropic and Meta, where experimental environments were mistakenly configured to allow models access to the outside world; Anthropic's agents reached real systems not part of the test, and Meta confirmed its model exploited a vulnerability in a third-party service after gaining internet access due to misconfiguration. Calcalist spoke with Michael Bargury, co-founder and CTO of AI security company Zenity, and Moshe Karako, Chief Technology Officer of NTT Israel, about the risks. Bargury explained that AI labs added cyber capabilities training, where models are given tasks like exploiting vulnerabilities, but in OpenAI's case, some tasks lacked necessary resources, such as the software itself, causing models to hit a wall and search for other ways to progress. The agents had access to JFrog's Artifactory, a "warehouse" for development files, which allowed them to find alternative paths to complete their tasks.