Securing AI Assisted Digital Evidence

AI-assisted digital forensics creates an assurance layer that conventional artefact-preservation controls were not designed to evaluate. Cryptographic hashing remains essential for detecting modification and is particularly salient under the current Indian electronic-evidence framework. In this study, SHA-256 detected all five controlled artefact modifications (5/5; 100% EIDR). The tested simple indirect prompt injection did not materially manipulate the AI-assisted forensic interpretation in any of ten fresh-session trials, and the protected condition likewise produced 0/10 successful injections. The negative result is retained as such: it does not establish universal robustness, and it prevents attributing an attack-prevention effect to the protected prompt. The protected workflow nevertheless improved the explicitness of evidence-to-finding mapping and classification of supported, partially supported and unsupported propositions. These observations support AIDEIF’s broader emphasis on processing traceability, provenance and validation even where a tested attack does not succeed. AIPCC complements conventional chain-of-custody records by capturing material AI-processing provenance; it does not replace statutory certificate requirements, forensic standards or judicial assessment. The central contribution is therefore a layered assurance model that preserves the distinction between the integrity of the evidentiary artefact and the reliability of the analytical process applied to it. Further multi-model and operational testing is required before making broader claims about attack resistance or the legal treatment of AI-assisted forensic conclusions.

[1] M. Scanlon, F. Breitinger, C. Hargreaves, J.-N. Hilgert and J. Sheppard, “ChatGPT for digital forensic investigation: The good, the bad, and the unknown,” Forensic Science International: Digital Investigation, vol. 46, Suppl., Art. no. 301609, 2023, doi: 10.1016/j.fsidi.2023.301609. [2] M. Chernyshev, Z. A. Baig, N. Syed, R. Doss and M. Shore, “Large language models in digital forensics: capabilities, challenges and future directions,” Forensic Science International: Digital Investigation, vol. 56, Art. no. 302043, 2026, doi: 10.1016/j.fsidi.2025.302043. [3] ISO/IEC 27037:2012, Information technology—Security techniques—Guidelines for identification, collection, acquisition and preservation of digital evidence, ISO, 2012. [4] Scientific Working Group on Digital Evidence, Best Practices for Digital Evidence Collection, SWGDE 18-F-002-2.0, Version 2.0, Nov. 20, 2025. [5] OWASP GenAI Security Project, OWASP Top 10 for LLM Applications 2025, 2024/2025 edition. [6] OWASP GenAI Security Project, “LLM01:2025 Prompt Injection,” 2025. [7] OWASP GenAI Security Project, “LLM08:2025 Vector and Embedding Weaknesses,” 2025. [8] K. Greshake, S. Abdelnabi, S. Mishra, C. Endres, T. Holz and M. Fritz, “Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection,” in Proc. 16th ACM Workshop on Artificial Intelligence and Security (AISec ’23), 2023, pp. 79–90, doi: 10.1145/3605764.3623985. [9] L. Huang et al., “A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions,” ACM Transactions on Information Systems, vol. 43, no. 2, Art. 42, pp. 1–55, 2025, doi: 10.1145/3703155. [10] X. Tan, H. Luan, M. Luo, X. Sun, P. Chen and J. Dai, “RevPRAG: Revealing Poisoning Attacks in Retrieval-Augmented Generation through LLM Activation Analysis,” in Findings of the Association for Computational Linguistics: EMNLP 2025, pp. 12999–13011, 2025, doi: 10.18653/v1/2025.findings-emnlp.698.