How researchers attack, monitor and constrain autonomous LLM agents that click, type and call tools in real environments, from prompt injection and quiet data leaks to runtime monitors and enforceable safety policies.
Small-plane cockpit: attitude indicator, compass and gauges against clouds · Photo: Rinald Rolle
A 2026 test of 12 AI models doing finance, legal and consulting work found that swapped or attacker-supplied scripts derailed most of them.