Uloženo v:
| Hlavní autoři: | , |
|---|---|
| Médium: | Recurso digital |
| Jazyk: | |
| Vydáno: |
Zenodo
2025
|
| On-line přístup: | https://doi.org/10.5281/zenodo.17601839 |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
Obsah:
- <p>Generative artificial intelligence (GenAI) is undergoing rapid expansion, driven by digital transformation across a wide range of sectors, with its applications expected to continue to grow. While GenAI offers significant benefits, its integration also introduces new attack surfaces and security vulnerabilities. In particular, Large Language Models (LLMs) pose new and unique security threats, including prompt injection, data and model poisoning, sensitive information disclosure, and hallucinations to name but a few. Ensuring the confidentiality, integrity, and availability (CIA) of data is fundamental to the development of secure GenAI systems. Among these, exacerbated by the proliferation of misinformation, disinformation, and deceptive content, threats to data integrity stand out as particularly critical. This issue is significantly compounded by the rise of multimodal models, which introduce further complexity to attack vectors by processing combined text, image and audio inputs.</p> <p>Although information manipulation has long existed as a feature of human behaviour, the scale, speed, and sophistication of GenAI powered attacks have now exceeded traditional detection capacities. Consequently, ensuring ethical and trustworthy AI development in critical sectors requires prioritising data integrity as a central concern.</p> <p>From the outset of its deployment, it became evident that GenAI presents novel security vulnerabilities, as indicated in scientific literature. Beyond acknowledging these problems, the systematic analysis of real-world attacks and their methodologies has become essential for developing effective countermeasures. </p> <p>The aim of the establishment of a GenAI incident database is to systematically document such occurrences, provide preventative measures through comprehensive incident analysis, and define attack taxonomies. This concept draws from proven methodologies in aviation safety and cybersecurity incident management. These databases serve a crucial role in education and training, enhancing awareness regarding prospective attacks and the consequences of AI related harm. This approach underscores critical ethical concerns in artificial intelligence, such as improper use, diverse biases, and inequitable algorithm performance, while concurrently offering pragmatic directions for the ethical advancement of GenAI systems. </p> <p>Both general purpose and domain specific databases have been established for documenting and analysing GenAI related incidents, with proposed taxonomies. Such databases serve to enhance transparency and foster collaboration among policymakers, researchers, and industry stakeholders alike. </p>