Signal A Brazilian bank deploys red agents alongside human experts for adversarial testing

Summary

Deloitte cites red teaming as one area where cyber teams are using AI to their advantage, stress-testing AI systems by simulating attacks before adversaries can exploit weaknesses. Brazilian financial services firm Itau Unibanco has recruited agents for these exercises. It combines human experts with AI test agents deployed across the company. These red agents follow an iterative process to find and mitigate risks such as ethical problems, bias and inappropriate content. Roberto Frossard, head of emerging technologies, says that in a regulated industry trust is the top concern, so the firm spends much time testing, retesting and simulating ways to break its models. The report adds that adversarial training, which exposes models to inputs designed to fool them, is another AI-native defence.