Topic

AI Misalignment

Gaps between a user's stated goal and the methods a model chooses, including spontaneous exploitation.

Current clusters

product1 distinct publisher

OpenAI says disclosure speed hinges on whether an incident looks like a traditional security breach

OpenAI published the Hugging Face incident the next day and treated a German wiki's agent edits as already-covered ground, which tells anyone running a public write surface what actually triggers a vendor advisory.

Publishers:engadget.com

Reality

Evidence48
Adoption56
Hype gap+14
Incentives78

build2 distinct publishers

A missing ownership check on cancel turned one gym member's assistant into an intruder

The booking software's cancel endpoint never checked who owned the reservation, so an assistant acting for one member erased another member's place and could not restore it. That exposure belongs to the product team.

Publishers:abc.net.auaikido.dev

Reality

Evidence74
Adoption55