Unpopular Opinion: Real Jailbreaking Is Mostly Dead in 2026... It’s Now Just Advanced Prompt Engineering
I’ll be direct.
In 2026, most things labeled as “jailbreaks” are no longer real jailbreaks. They’re advanced prompt engineering.
A true jailbreak forces a model to do something it was structurally designed to refuse, even with good prompting. What we mostly see today are long, cleverly written prompts using roles, hypotheticals, iteration, and framing. These techniques work well, but they’re not bypassing the model’s core restrictions... they’re working around them through better communication.
The old one-shot DAN-style jailbreaks that actually broke the rules have become rare on current models. Safety systems have improved significantly.
Personally, I see this as a positive shift. It means models are getting smarter, and we’re learning to prompt them more effectively instead of relying on tricks.
Where do you draw the line between advanced prompt engineering and actual jailbreaking in 2026?
Do you think real jailbreaks still exist, or has everything become sophisticated prompting?