Lethal Injection

Hijacking LLM-Driven Malware Triage from Inside the Binary

Master Thesis (2026)
Author(s)

E.J.E. Handgraaf (TU Delft - Electrical Engineering, Mathematics and Computer Science)

Contributor(s)

H.J. Griffioen – Mentor (TU Delft - Electrical Engineering, Mathematics and Computer Science)

J. Blaser – Mentor (Vrije Universiteit Amsterdam)

S. Proksch – Graduation committee member (TU Delft - Electrical Engineering, Mathematics and Computer Science)

Faculty
Electrical Engineering, Mathematics and Computer Science
More Info
expand_more
Publication Year
2026
Language
English
Graduation Date
10-09-2026
Awarding Institution
Delft University of Technology
Programme
Computer Science
Faculty
Electrical Engineering, Mathematics and Computer Science
Page Views
69
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract

LLM agents are moving into malware triage, reading binaries through tools such as GhidraMCP. Letting an agent return the verdict with no human in the loop is a short step away. The pipeline cannot separate the content it analyses from instructions to follow, and that content is authored by the very party the verdict aims to catch. Such indirect prompt injection has already appeared in the wild, crude and unsuccessful; we ask what a well-formed can achieve.

To answer that question, we present Lethal Injection: a large-scale, in-depth analysis over the attack surface of indirect prompt injection in the context of fully automated agentic malware triage. We sweep models, payloads, and deployment configurations in a single-prompt setting, then run the attack end to end, feeding real malicious binaries through a GhidraMCP-enabled agent. Model choice is a security variable, though neither capability nor model size predicts robustness. One payload we tested flips every sample on seven of eleven models, across four model families. What a payload claims matters more than how it is written. The payloads that succeed assert a premise the model cannot check and let it reason correctly from a poisoned start. Character-level encoding and iterative augmentation add little on the models we sweep for them. Injections compiled into the binary are recovered by the agent’s retrieval tools and flip verdicts on real malware — on every sample for one model, on none for another. In one run an injected string steered a state-changing tool call. A single injection-aware line in the system prompt cuts bypasses sharply on the two most susceptible models, at the cost of more false positives on benign-but-suspicious code. Across four model families an injection turns a malicious sample benign, so a benign verdict from an undefended pipeline is not evidence about the binary.

Files

Report.pdf
(pdf | 3.22 Mb)
License info not available