2026-W20 · +0.4
This week’s dossier points to modestly higher concern about AI capabilities and acceleration, mostly because of new work on jailbreaks, agent red-teaming, and automated evaluation. There is limited additional evidence on autonomy: some papers argue that oversight and automated alignment may be harder than expected, but they do not show real-world loss of control. No new severe real-world AI incident is documented in the provided sources. Governance remains mixed, with concerns about reproducibility of safety claims but only tentative signs of possible new oversight.
A conventional composite index, not a scientific probability of extinction.