Browse this collection and consult each publication for its evidence, sources and stated limitations.

ODA3-2026-11-INS-120
5 Nov 2026
The text an AI generates, the version a person approves and the message a recipient receives may differ. A defensible account needs evidence across those hand-offs, and a communication record…
Read publication →
ODA3-2026-10-INS-097
14 Oct 2026
An analysis of 857 Chinese model releases, the limits of safety disclosure, and four evidence checks for enterprise AI approval.
Read publication →
ODA3-2026-10-INS-095
13 Oct 2026
AI-found attack chains are leads, not findings. Using Rubrik Code Guardian as a case, ODA3 sets out the evidence a red-teaming pilot must actually produce.
Read publication →