At its core, they mentioned, “AI red teaming strives to push past model-level security benchmarks by emulating real-world assaults towards end-to-end programs. Nevertheless, there are numerous open questions on how red teaming operations ought to be performed and a wholesome dose of skepticism in regards to the efficacy of present AI red teaming efforts.”
The paper famous that, when it was fashioned in 2018, the Microsoft AI Red Team (AIRT) centered totally on figuring out conventional safety vulnerabilities and evasion assaults towards classical ML fashions. “Since then,” it mentioned, “each the scope and scale of AI red teaming at Microsoft have expanded considerably in response to 2 main developments.”
The primary, it mentioned, is that AI has develop into extra subtle, and the second is that Microsoft’s current investments in AI have resulted within the growth of many extra merchandise that require red teaming. “This enhance in quantity and the expanded scope of AI red teaming have rendered absolutely guide testing impractical, forcing us to scale up our operations with the assistance of automation,” the authors wrote.