Research on how open-source AI is defined, governed, documented and distributed before deployment is sparse, according to a review for the UK government published on 7 September [1].
The review, by researchers at the University of Greenwich, found academic work strong on downstream risks, such as detecting intrusions and making models robust to attack [1].
But it found the work "near-silent" on how models are trained, what data they use and how weights are distributed [1].
It found definitions of open-source AI inconsistent across governments, industry and academia, and says that inconsistency is itself a governance risk: "Where the definition stops, security responsibility disappears" [1].
What it measures
Published literature from 2020 to 2026. The academic search covered eight databases and screened 14,561 records, of which 43 met the criteria [1].
It also reviewed 172 reports from national cyber security authorities, standards bodies, international organisations and open-source groups, and analysed GitHub and Hugging Face [1].
It points to the Open-Source AI Definition 1.0 of October 2024, which requires training-data information, source code and model parameters to be released together, as the most comprehensive definition available [1].
What it does not tell you
What the UK government will do. The report says it does not represent government policy [1].
How secure open-source AI actually is. It reviews what has been studied, not the systems themselves [1].
Comments