SafeCover Automation Error Costs $12.5 Million
When unchecked AI met 'fail-open' defaults, SafeCover paid the price.
By midday on February 23, SafeCover Insurance’s claims floor was overwhelmed. The LiteLLM automation suite approved nearly $12.5 million in fraudulent payouts. No hack occurred. The system processed the claims without error-checking.
“They pushed this live without testing. We’re cleaning up their mess,” said one junior analyst.
JSON validation fields were set to “fail-open.” Any minor lag in database responses went unchecked. Scammers exploited the loophole immediately, sending malformed files and altered timestamps. LiteLLM accepted them as valid.
Over 2,100 claims were approved in hours. More than 3,400 internal red flags were ignored. Some claims were paid multiple times within the same hour.
By February 25, over 800 claims bypassed all system checks. The claims department switched to manual auditing, working long shifts to correct the errors.
Management’s attempt at a fully autonomous pipeline failed. The $12.5 million loss is listed as operational expense. There was no mitigation plan. No employee bonuses. No public explanation. The system remains suspended, and oversight has been reinstated.
