Alignment technique improved through 2026; the gap between capability and verifiable safety widened anyway
Surveys of the year describe genuine progress in RLHF, constitutional methods and mechanistic interpretability — alongside a continued widening of the distance between what models can do and what anyone can verify about them. Both statements are true, and holding them together is the honest position.
The temptation is to resolve the tension by picking a side: either technique is advancing, or safety is losing ground. The uncomfortable reading is that both hold because they measure different things. Technique improves linearly with effort; the verification burden grows with the space of behaviours a model can express, and that space has been growing faster.
Agentic deployment is what turned this from a theoretical concern into an operational one. A model answering questions has a bounded output surface that can be sampled. A model taking actions across tools and time has a behaviour space that cannot be enumerated, let alone tested, which is why reliability moved to the centre of the evaluation conversation this year.
The practical response visible in industry is a retreat to containment: sandboxing, egress control, human review at the boundary. That is a sound engineering instinct and an admission — you contain what you cannot verify. For H2 2026, the containment layer is where the real safety investment is going, whatever the research agenda says.
6G-AI — The Alignment Problem in 2026: Progress, Setbacks, and the Road Ahead → · AI Made — AI Safety in 2026: What the Research Actually Shows →