One metric for prompt injection and temporal blindness. LLM agents read control
instructions and ordinary data through the same channel, so they must decide whether to obey each
instruction.
The benchmark measures two symmetric failure modes: over-compliance (obeying illegitimate
signals — prompt injection) and under-compliance (ignoring legitimate ones). The
Signal-Response Correctness (SRC) metric ranges 0–1 per item;… See the full description on the dataset page:
https://huggingface.co/datasets/thamilvendhan/signalbench.
