Balkanization of execution security in AI agents

The growing adoption of artificial intelligence agents capable of operating autonomously —reading repositories, executing commands, calling APIs— has opened a new frontier in cybersecurity. However, research on the security of their execution layer is fragmented: studies on isolation in controlled environments, capability control, access policies, TOCTOU (time-of-check-to-time-of-use) attacks, vulnerabilities in protocols such as MCP, identity delegation, or static analysis of AI-generated code rarely cite each other. This disconnect creates critical gaps that no single work addresses comprehensively, such as the lack of common benchmarks to compare isolation architectures, the high failure rate (between 69% and 98%) of real-world denylists without isolation systems reconsidering them, or the fact that benign but out-of-scope actions occur in up to 17.1% of cases without control mechanisms mitigating them. These shortcomings are especially relevant for companies integrating AI for business into their processes, since a poorly protected agent can become an attack vector. At Q2BSTUDIO, as a custom software development firm, we combine our expertise in tailored applications with cybersecurity services to design robust AI agents. Additionally, we deploy these solutions on AWS and Azure cloud services and complement them with business intelligence services such as Power BI to monitor their behavior. Our approach integrates granular access controls, continuous auditing, and policy analysis, closing the gaps that academic literature has yet to resolve. Ultimately, the security of AI agents is not just a technical problem, but a governance challenge that demands collaboration between researchers and development professionals.