Indirect prompt injection is a growing threat for LLM-based systems that pull in untrusted data from external sources. Unlike direct injection, where a user tries to override instructions, indirect injection hides inside retrieved documents, tool outputs, or messages from other agents. The new arXiv paper introduces TRACER, a training method that uses trace regularization to reduce how often such hidden instructions override a model's intended behavior.
The paper's key claim is that defense robustness is channel-conditioned: the same attack does not behave identically across different delivery channels. That means a defense that works for retrieved web content may fail for tool outputs or agent messages. TRACER is designed to account for this by regularizing the model's response traces, encouraging consistent behavior regardless of how the injection arrives.
The abstract reports that under TRACER, the same attack succeeds only 18% of the time, but the sentence is cut off before the baseline comparison is given. Because the source is a single arXiv abstract, the full experimental setup, baseline numbers, and limitations are not yet available. The work points toward a more nuanced view of prompt-injection defense: instead of a single robustness score, defenses may need to be evaluated per channel.