Prompt injection is widely considered the dominant security risk for large-language-model applications and agents, according to a new systematization-of-knowledge paper posted on arXiv. The paper, titled "SoK: PIAtlas: Decomposing the Unbounded Prompt Injection Attack Space," argues that the field still lacks a coherent map of how these attacks are constructed.
The abstract notes that more than twenty public datasets describe prompt-injection attacks, yet no unified framework ties them together. To address this, the authors propose PIAtlas, a decomposition of the unbounded prompt-injection attack space.
Because the abstract is truncated, the full methodology and findings are not available from the announcement alone. The paper's contribution, as described, is to provide a structured way to understand and compare prompt-injection attacks across the many existing datasets.