Abstract

Agentic systems for biomedical discovery are overwhelmingly optimised to generate candidates and gather supporting evidence. The failure mode this creates is specific: an analysis that looks successful passes, and the errors that survive are precisely those whose apparent performance is good. We propose a three-agent framework in which a blinded attacker adversarially attacks each proposed biomarker before a rule-based arbiter may accept it, with all computation delegated to a deterministic tool layer so that verdicts are reproducible. To evaluate it we do not construct synthetic errors. We use four analysis errors that actually occurred during a spatial-transcriptomic study of antibody–drug conjugate (ADC) targets in small cell lung cancer, each documented at the time it was found. With the attacker disabled, two of four flawed proposals are accepted; with it enabled, none are, while the correct proposal is retained. A coverage analysis shows the errors are usually caught redundantly (2.25 attacks per error) but that negative-control synthesis is a single point of failure: removing it alone lets an error through, and it appears in every minimal sufficient attack set. The surviving specification yields a clinically consequential result: along a neuroendocrine axis built without the targets, DLL3 and TROP2 cross, so subtype drift does not erode an ADC target but exchanges one for another. Two caveats prevent overstatement: two of the four errors are also caught by confirmation alone, and a single-model test of blinding was inconclusive.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_014.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=c3RN0QmAdI

BibTex

@InProceedings{HonSun_Attack_MICCAISAT2026,
        author = { Hong, Sungman AND Chun, Harim AND Lee, Hanjoo AND Ahn, Sangjeong},
        title = { { Attack Your Own Findings: A Falsification Agent Framework for Biomedical Discovery, Benchmarked on Four Errors We Actually Made } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17263},
        month = {pending},
        page = {pending}
}


back to top