NTBench: near-target prompt injection against correct-looking task outcomes

Evidence record

NTBench: near-target prompt injection against correct-looking task outcomes

Read the original source ↗

Claim-level citations (2)

• supportsNear-target prompt injections preserve the main task while changing a critical detail: NTBench checks deviations from a task-specific reference and excludes mistakes already present without injection, separating maliciously induced near-target changes from baseline agent error.

  Abstract, reference-derived targets and exclusion of uninduced deviations

• supportsNear-target prompt injections preserve the main task while changing a critical detail: The paper compares near-target objectives with fixed-target prompt injections across seven LLMs, ten datasets and two agent benchmarks, reporting larger measured deviation rates for the close-to-correct attack targets.

  Abstract; attack-target construction and controlled evaluation