Brendan Dolan-Gavitt

@moyix · Twitter · 2026-04-18 15:05 UTC

GPT-5.4 got a successful exploit (waaay harder than the others, requiring multiple strategic nudges - though not hints since I didn't know the solution). Sadly the bug it found and exploited is one that is publicly known so I'm thinking of asking it to try again...

Brendan Dolan-Gavitt

@moyix · Twitter · 2026-04-15 16:41 UTC

Here we go again...