EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Telemetry Ablation for Post-Incident Prompt-Injection Reconstruction: Evidence Sufficiency and Human Source Attribution in an LLM–Integrated Application

Vedran Dakić, Zlatan Morić, Zdravko Kunić, Damir Regvart

Journal of Cybersecurity and Privacy · 06.11.2026 · 6 (6) · 195

DOI: 10.3390/jcp6060195 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Cybersecurity and Privacy, online erschienen am 06.11.2026. Verfasst von Vedran Dakić und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Prompt injection in large language model (LLM)–integrated applications makes forensic readiness necessary when prevention fails. We model execution as a typed provenance multigraph and derive evidence packages by deterministic, idempotent projection from a canonical trace, isolating telemetry effects from model nondeterminism. The model establishes monotonicity of attainable reconstruction and conditional non-identifiability when retained context collapses distinct source-labeled inputs. Seven investigators reconstructed 168 sy1nthetic incidents, yielding 1176 responses across seven assignment arms representing six unique evidence sets; one ablation duplicated L2 and is treated as a consistency arm. The minimum forensic logging profile (MFLP) decision rule, frozen before scoring but after investigator response collection had begun, returned ABSTAIN: at complete telemetry, one-sided 95% lower bounds for frozen exact object attribution, path coverage, and conditional detection recall were 0.123, 0.646, and 0.880, below the 0.90 adequacy floor. With complete telemetry, investigators’ post-incident classification (detection) recall was 0.921 and channel-level attribution was 0.688, whereas exact object attribution was 0.176 (95% confidence interval: 0.113–0.246); the exact score is partly depressed by an object-label mismatch in the instrument, which capped attainable exact attribution at 0.832. Retrieval provenance and tool telemetry contributed differently; we did not establish an incremental value for the environment/network. No empirical minimum profile is identified. The framework is model-agnostic; estimates apply to the evaluated application, corpus, model, and investigator sample.

Zitieren

APA

Dakić, V., Morić, Z., Kunić, Z., & Regvart, D. (2026). Telemetry Ablation for Post-Incident Prompt-Injection Reconstruction: Evidence Sufficiency and Human Source Attribution in an LLM–Integrated Application. Journal of Cybersecurity and Privacy, 6(6), 195. https://doi.org/10.3390/jcp6060195

Vancouver

Dakić V, Morić Z, Kunić Z, Regvart D. Telemetry Ablation for Post-Incident Prompt-Injection Reconstruction: Evidence Sufficiency and Human Source Attribution in an LLM–Integrated Application. Journal of Cybersecurity and Privacy. 2026;6(6):195. doi:10.3390/jcp6060195

BibTeX
@article{Daki2026,
  author = {Dakić, Vedran and Morić, Zlatan and Kunić, Zdravko and Regvart, Damir},
  title = {Telemetry Ablation for Post-Incident Prompt-Injection Reconstruction: Evidence Sufficiency and Human Source Attribution in an LLM–Integrated Application},
  journal = {Journal of Cybersecurity and Privacy},
  year = {2026},
  volume = {6},
  number = {6},
  pages = {195},
  doi = {10.3390/jcp6060195}
}
Kostenlos starten →