EconSafeBench shows LLM agents break financial rules unprompted
The gap between refusing explicit instructions and violating rules under goal pressure means safety reviews based on red-teaming prompts will not catch real deployment risks in trading or pricing agents. Teams building agentic systems for finance, hiring or procurement should add target-only test conditions and compliance checks at escalating goal levels to their evaluation suites. The tool-call-based scoring approach is reusable beyond the financial cases in the benchmark.