Beyond Instruction Override: Benchmarking Indirect Prompt Injection with Task-Tampering Attacks
Source check: Source checked. Not independent scientific confirmation.
Finding: A benchmark of 90 validated cases across three tool-use environments evaluated whether existing defenses covered attacks that tamper with decisions within a user’s task. Detection-and-removal reduced objective-targeting success more than observation- and state-targeting success. This is a defensive benchmark result; attack construction details are omitted.
Technology: Tool-using language models and detection/removal and tool-policy defenses
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Tool-using language models and detection/removal and tool-policy defenses
- Test and result
- In 90 benchmark cases across three tool-use environments, detection-and-removal reduced objective-targeting success to 2.4%, while observation-targeting and state-targeting success remained 9.5% and 12.1%.
- Limitation / still unknown
- Results are benchmark-specific and do not establish operational deployment or general defense rates. Attack construction details are intentionally omitted.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2611.07640
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.