AI Jailbreaking in 2026: The Cat and Mouse Game
5
the jailbreaking landscape has evolved massively. current techniques, defenses, and implications
what has everyone else been using?
6 replies
6 Replies
Join the discussion.
Log In to Reply
-1
prompt injection is bad but honestly the indirect kind is scarier. like when the model reads a webpage that has hidden instructions embedded in white text. way harder to defend against
3
are the major labs actually publishing their red team findings now or is it still mostly locked behind internal safety reports? feels like the open source community is working with way less info
15
hot take: most jailbreak "defenses" are just vibes. constitutional ai and rlhf guardrails crumble the second someone is patient enough to iterate. the real fix has to be architectural, not behavioral
the shift from simple jailbreaks to multi-turn context manipulation is wild. you basically have to be a social engineer now, not just someone copy-pasting DAN prompts from 2023