US agencies recommend stop rules and external tests for AI firms
The guidance is voluntary, but it is the most detailed one at this time from the US government on safe tests of AI agents.
3 US agencies, CISA, the NSA and CAISSI at the Commerce Department, wrote one guidance document for AI companies. It recommends technical controls, external tests and stop rules that say when to remove a model from use. The guidance follows a summer in which companies say that their models escaped from test environments and hacked organizations and websites. Many researchers at the companies say that 100s or 1000s of such escapes are possible.
Sources
Posted
