A new framework called Palette lets AI companies flip a switch and loosen specific safety refusals for approved users, without retraining the whole model.
Researchers behind the method say current safety alignment is one-size-fits-all: the same refusal policy applies to every user and every context, even when a request is legitimate for an authorized professional. Palette works by identifying a "refusal direction" inside the model through multi-objective search, then baking that adjustment into the model with lightweight adaptation rather than a full realignment. It also supports modular composition, meaning domain-specific safety controls can be trained separately and merged later, so a company could add new authorized use cases on demand. The team tested it across four safety benchmarks and multiple model variants, including both language and vision-language models, and reported that it preserved general safety and usefulness outside the authorized domains.
The pitch is really about access control, not just safety. Instead of a single refusal threshold for everyone, Palette treats permission like a dial that can be turned per domain and per user group, which is a meaningfully different architecture than blanket jailbreak patches or prompt-level steering.
That reframing only works if whoever controls the dial can be trusted to decide who counts as "authorized," and the paper never says how that verification would happen in practice.