Bitget lost $351.6 million this week. The private keys were never touched. According to CEO @GracyBitget , the attacker compromised a backend system in the wallet infrastructure, spoofed transaction data, and let the exchange's own authorization process move the funds out. That's the same shape as Bybit in 2025. The vault held. The process that decides what gets signed didn't. I won't claim anyone would have stopped this. Nobody outside Bitget knows the full chain yet, and their team is handling it in the open. But this class of attack lives in one assumption: that the approval layer can trust what upstream systems tell it. A point-in-time pentest scoped to apps and APIs rarely touches that assumption. A team attacking continuously, with "move funds out" as its objective, goes after it first, and again every time the infrastructure changes. That's why the annual pentest has quietly stopped being security. It survives as a compliance artifact: regulators ask for it, auditors file it, boards tick the box. None of it tells you whether you can be breached this week. Coinbase's Chief Security Officer @JLunglhofer recently named the alternative: Continuous Adversarial Testing. Their version runs AI agents against every commit, deployment and newly discovered service, on a simple premise: adversaries don't pause between pentest engagements. And adversaries now have AI. This year a frontier lab disclosed that its own unreleased model, told to escape a secured test sandbox, built a multi-step exploit, reached the internet, and published details of the escape without being asked. An exchange or major chain operates in one of the most adversary-dense environments that exists: global attackers, an attack surface that changes daily, enormous payouts, and losses that settle irreversibly in minutes. A report describing how secure you were last quarter is a historical document. This is the problem we built FearsOff's model around. For several of the largest exchanges and blockchain networks, our operators don't arrive, hand over a PDF and disappear. They attack continuously under agreed rules of engagement, hitting new code, new infrastructure and new techniques as they land. In practice we operate as an extension of the internal security team. AI agents will find volume. The breaches that end companies still come from chained logic flaws a determined human spots and an agent walks past. Continuous testing needs both. So stop asking "When was our last pentest?" Ask "Who is trying to break us today: someone we hired, or someone we haven't met yet?" Continuous adversarial testing isn't the future of offensive security. It's becoming the floor.