The 10 most expensive software failures and what they share
From Knight Capital to Boeing 737 MAX, Therac-25 to CrowdStrike: examining the 10 costliest software failures in history and their shared root cause.
Looking across software history's ten costliest incidents reveals a striking pattern: almost none were attacks. Except for NotPetya, every case on the list stemmed from a silent, scalable failure generated by an organization's own system. Knight Capital lost $440 million in 45 minutes while every system reported itself healthy — the root cause was a feature flag tied to code retired years earlier but never removed.
The list spans industries with a recurring shape. Boeing's MCAS trusted a single faulty sensor, contributing to 346 deaths. In the UK's Post Office Horizon scandal, phantom shortfalls generated by accounting software led to hundreds of wrongful convictions. Zillow's pricing algorithm didn't misprice one house — it systematically overpaid across an entire portfolio. Ariane 5, Mars Climate Orbiter, the 1990 AT&T outage, and Therac-25 each show how dead code, unit mismatches, and race conditions turn catastrophic once unverified. CrowdStrike's 2024 outage is the most recent proof of how a single trusted update can propagate worldwide almost instantly.
What unites these cases is that systems kept reporting green while errors compounded silently, and one change spread across an entire network before humans could intervene — no adversary required. For engineers, the real takeaway is that this exact risk is poised to resurface, at far greater scale, with autonomous AI agents: systems that act fast and consistently without human oversight can produce expensive failures even while appearing correct.