LLMs that operate over external data—such as documents or tool returns—face a tricky situation: a single piece of content can interleave legitimate directives with adversarial prompt injections. The model must obey the former and ignore the latter, but doing so reliably is difficult when both appear in the same input.

The arXiv paper 2610.22874 proposes a framework called "Identify, Classify, Act" to address this per-instruction tension. Rather than treating the entire input as a single block, the approach is designed to evaluate each directive individually, deciding which ones deserve compliance and which should be rejected.

The abstract emphasizes "per-instruction controlled compliance," suggesting that the model's response should be conditioned on the nature of each separate instruction. This is a step toward more robust handling of untrusted external content, though the abstract provides limited detail on the underlying mechanism.