Testing beyond one internal team

Microsoft announced its External Red Team Alliance, EXTRA, on 27 July. The programme extends collaboration beyond the company’s own AI Red Team and supports academic research through unrestricted gifts. Microsoft said the funded assessment work spans six continents.

Language and domain knowledge matter

The announcement argues that important failure modes can depend on specialist expertise, local context and languages that a single internal team does not fully represent. Its remit includes misuse, security and alignment-related assessment. EXTRA is a research and collaboration initiative. The useful outcome will be evidence from the assessments it supports, rather than treating the number of participating regions as a measurement of how safe a model has become.

Sources & further reading

Enhancing AI security through global AI red teaming

Published by Capvise Media. Editorial standards · Corrections