Security practices in AI development

Security practices matter more than tools alone: The paper finds that how companies implement safety measures (like red teaming and alignment procedures) is more important for building trust than the capabilities of the tools themselves, suggesting that process and transparency are critical to credible safety claims.

There's a significant gap between what safety tools promise and what they deliver: The research identifies a fundamental mismatch between the capabilities of current security tools and the safety guarantees they're meant to provide, meaning companies cannot fully guarantee AI safety despite using these tools.

Current practices lack diversity and meaningful participation: The paper identifies "shortcomings in diversity and participation" in how AI security practices are developed and implemented, suggesting that safety decisions are being made by too narrow a group without sufficient input from affected communities.

Safety claims are about managing perception, not absolute guarantees: The authors describe AI security practices as part of "securitization processes" designed to make AI development seem trustworthy enough to proceed commercially, rather than actually guaranteeing safety—meaning trustworthiness can only be "imperfectly tested" rather than proven.

Current practices need structural improvements: The paper concludes with recommendations for improving AI security practices, implying that the status quo is insufficient for ensuring responsible AI development and deployment.