"There is a real risk of people enthusiastically copy-pasting documents or classified information and that information then being used to train the model." [classified information]
The quote names the specific mechanism: employees unknowingly feed sensitive data into AI training pipelines via normal workflows, creating a systemic vulnerability that governance frameworks cannot prevent because the risk is embedded in user behavior, not just system design.
Reasoning from this article
The article frames this as a governance failure, not a technical one. Matthew Hodgson (Element CEO) and Bill Conner (Jitterbit CEO) both emphasize that companies 'have very little idea' of data governance and assume cloud hosting (AWS/Azure) provides safety when it does not. The Slack and Microsoft incidents are presented as evidence that even major platforms cannot prevent data leakage. This suggests a structural mismatch: AI adoption is outpacing the organizational capability to control data flows, creating a persistent vulnerability class.