Researcher Argues Transformers Cannot Separate Instructions From Context
Developer gerardsans argues that Transformers inherently cannot distinguish instructions from observations within a context, undermining the boundaries needed for true agents and serving as a major cause of jailbreak attacks.
2026-11-21 ~ 2026-11-21 · 2 related posts
- Dev argues transformers can't separate instructions from context, enabling jailbreaks — gerardsans · 2026-11-21
- Transformer can't distinguish instructions from observations, undermining real agents, researcher argues — gerardsans · 2026-11-21