OnlyTech.boo

Silent Killers - The Therac-25 Radiation Disaster

April 6, 2026$25k estimated cost

Context

In the early 1980s, the Therac-25 was marketed as a cutting-edge medical breakthrough—a radiation therapy machine controlled almost entirely by software. It was faster, more efficient, and required fewer hardware safety mechanisms than its predecessors. It was also dangerously flawed.

What Happened

Between 1985 and 1987, in hospitals across North America, something unthinkable began to happen. Patients undergoing routine radiation therapy started reporting intense burning sensations - far beyond anything expected. Operators saw nothing unusual. The machine displayed normal readings. No alarms. No warnings. But inside the system, a hidden race condition was being triggered. Under specific sequences of rapid input by operators, the machine’s software would enter an inconsistent state-silently disabling safety checks and firing a concentrated, lethal dose of radiation directly into patients. Massive overdoses-hundreds of times the intended level-were delivered in seconds. Some patients died. Others suffered catastrophic injuries. And for a long time, no one could explain why.

Root Cause

Undetected race conditions in concurrent software Complete reliance on software without hardware safety interlocks Poor testing and dismissal of early warning signs

Impact

Multiple patient deaths and severe injuries Lawsuits and widespread fear of software-controlled medical systems One of the most infamous failures in the history of engineering

Fix

The system was eventually redesigned with: - Hardware safety interlocks reinstated - Extensive software validation and testing - Stronger regulatory oversight in medical devices

Lessons Learned

  • Software failures can be physically deadly in real-world systems
  • Race conditions are subtle but catastrophic
  • User interfaces can dangerously misrepresent reality
  • Ignoring early anomalies leads to disaster

Prevention

  • Always implement hardware fail-safes in safety-critical systems
  • Conduct independent and adversarial testing
  • Design systems to default to safe states on failure
  • Ensure transparency between system state and operator feedback

Similar incidents

THE DAY MICROSOFT WENT DARK

0 upvotes
$1200000k

Inside the Microsoft 365 Global Outage - A Discovery Documentary Microsoft 365 is more than an office productivity suite. It is a deeply interconnected cloud ecosystem supporting email, enterprise communication, file collaboration, identity-dependent workflows, security operations, business applications and AI-powered productivity. For millions of organizations, Outlook and Exchange Online form the backbone of daily operations. Teams, SharePoint, Copilot, Microsoft Graph and numerous third-party applications depend on Microsoft infrastructure and, in certain workflows, Exchange Online itself. This creates an important dependency: a failure in one critical service can affect several seemingly independent products. On November 12, 2026, Microsoft began experiencing a major service disruption that affected Exchange Online and progressively disrupted dependent Microsoft 365 functionality across multiple regions. The consequences extended beyond lost working hours. Financial transactions were interrupted, airlines experienced email-support problems, financial institutions faced operational disruption, and companies were forced to reschedule corporate activities. Experts in Australia warned that the economic damage there alone could reach billions of dollars, although the government said the actual impact remained unquantified. The final economic cost will depend on how much business activity was permanently lost rather than merely delayed, and how long recovery and operational disruption continued. The incident raises a fundamental question about modern cloud infrastructure: How can a platform operating at Microsoft's scale lose so much service capacity, and why did recovery prove so difficult?

PocketOS operated as a SaaS platform for car rental businesses, running on cloud infrastructure with shared storage volumes across staging and production. An AI coding agent inside Cursor, powered by a model from Anthropic, was granted execution capabilities within this environment. The system served real customers with live transactional data. A small engineering team managed infrastructure, application logic, and deployments. Stakeholders included rental operators, end users, developers, and infrastructure providers such as Railway.

Comments

Oldest first.

Loading comments…