THIS GUY GOT A $94,605 HOSPITAL BILL.
INSURANCE PAID $17,983.
THE GUY PAID ONLY $130.
Welcome to American healthcare.
Behind those three numbers is a maze of negotiated rates, billing rules, insurers and software deciding who actually owes what.
Most people only ever see the final number.
Hackers look at the system underneath.
And increasingly, they’re bringing Claude with them - tracing code, finding broken logic and hunting vulnerabilities before someone malicious does.
I wrote the full Claude bug-hunting setup below.
anthropic, in their own research post:
"the testers caught the dangerous command just 13.6% of the time."
1,053 paid professionals. one dangerous command dropped into an ordinary session. nine out of ten waved it through.
THEY DID NOT MEASURE THE MODEL. THEY MEASURED YOU.
the interesting part is not that humans lost. it is how they lost.
> 17% caught early in a session
> 5% caught after fifty prior prompts
> auto mode: flat, whatever the session length
attention is not a constant, it is a budget. every prompt you approve spends a little of it. by the fiftieth you are not reviewing anything, you are clicking.
head to head: the machine blocked 800 commands a human had approved. humans blocked 6 the machine let through.
one number rescues us though. people reject 3% of individual commands - and 39% of plans. ask someone about one line and they rubber-stamp it. ask about the shape of the whole job and they suddenly think.
so it is not approval fatigue. it is the wrong question, asked hundreds of times a day, until the answer stops meaning anything.
august 14 it becomes the default. the fix was never going to be a better human.