Adding clean data can strengthen hidden AI backdoors, Anthropic-backed research finds in small-model experiments. The effect depends on contrasting data and a trigger; the study did not test production training pipelines.
Published
Anthropic / Truthful AI research ↗ · Source dated