When Agents Attack: How Prelude's Engineering Team Fights Back

Prelude's engineering team has observed a surge in automated fraud attempts as AI agents from OpenAI, Anthropic, and Meta breached real organizations during security evaluations. The team notes that agents are not smarter than humans but faster, enabling attackers to run multiple parallel probes. Prelude is adapting its fraud detection by evolving SMS pumping detection, using network path as a signal, and deploying internal agents to uncover undetected fraud, with plans to offer similar capabilities to customers.

Originally published on the Prelude engineering blog https://prelude.so/blog/fighting-agents-with-agents . When OpenAI, Anthropic, and Meta each disclosed within weeks of each other that their own models had breached real organizations during security evaluations, I had one question: what does this actually mean for fraud detection? The short answer: agents haven't gotten smarter than humans. They've gotten faster. A single attacker can now run 10, 20, or 30 agents probing your system in parallel. We've seen automated behavior across customer applications increase exponentially. And the techniques aren't exotic: weak passwords, unprotected endpoints, basic SQL injection. The door was always there. Agents just find it faster. In this post, I cover what Prelude actually sees when an agent hits a login or OTP flow, how SMS pumping detection is evolving as agents learn to probe thresholds, why we've shifted toward network path as a signal that agents can't easily fake, and how Prelude is already using its own internal agents to find fraud we haven't caught yet. The part I'm most excited to share: what we're building next to put that same capability directly in customers' hands.