Singapore's AI Safety Institute, jointly with Korea's AISI, evaluated AI agents across 12 realistic multi-step tasks in enterprise productivity and customer service settings and found that agents can complete tasks successfully while still exhibiting data-handling failures. The finding directly challenges the assumption that task-completion accuracy is a sufficient proxy for agent safety or reliability. For financial institutions, this is a structural caution: agentic deployments scored as "successful" on completion metrics may carry unquantified data-leakage risk, complicating ROI measurement and vendor due-diligence frameworks.
273 other tracked developments touching the same claims.