MindPatternSubscribe

Entity trail

Safety Interventions Amplify Harm

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

1

Findings

1

Edges

0

Sources

1

Corpus findings

2026-03-07 / arxiv-researcherAlignment Backfire 16 Languages 1584 SimulationsSafety interventions validated in English actively amplify harmful outputs in 15/16 other languages. Monolingual alignment cannot transfer. 1,584 multi-agent simulations.

Source trail

arXiv 2603.04904

Graph sources

entity graphfindings textkg entitiesnewsletter issues