Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{206104,
        author = {Apurva Kolhe},
        title = {Reverse Engineering in Artificial Intelligence Systems: Techniques, Challenges, Security Risks, Interpretability, and Future Directions},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {2},
        pages = {417-426},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=206104},
        abstract = {Reverse engineering of artificial intelligence (AI) systems has emerged as a critical research frontier that simultaneously threatens intellectual property and privacy while enabling auditing, interpretability, and safety verification. This survey consolidates a decade of work spanning model extraction, adversarial perturbation, gradient inversion, membership inference, and the structural analysis of neural networks and transformers. We organize the field along two orthogonal axes—the adversary’s access model (black-box vs. white-box) and the recovery target (functionality, parameters, training data, or architecture)—and show how these axes unify otherwise disparate attack families. Particular attention is given to large language models (LLMs), where prompt injection, jailbreak analysis, and partial weight reconstruction have demonstrated that even commercial API-gated models leak exploitable structure. We connect offensive reverse engineering to its defensive and scientific counterpart, mechanistic interpretability and explainable AI (XAI), arguing that both pursue the same underlying objective: converting opaque parameter tensors into human-legible mechanisms. A comparative analysis of attack–defense pairs, a historical timeline, and a discussion of open-source versus proprietary vulnerability surfaces are provided, followed by enterprise security implications and a forward look toward reverse engineering of autonomous agents and AGI-scale systems. We deliberately mark uncertain citations as placeholders rather than fabricate references, and we close with suggested figures, an implementation architecture, and derived project ideas.},
        keywords = {Reverse engineering, model extraction, adversarial machine learning, membership inference, gradient inversion, large language models, prompt injection, mechanistic interpretability, explainable AI, AI security, transformers, autonomous agents.},
        month = {July},
        }