UK researchers find that random control steering vectors can suppress evaluation awareness in AI models just as effectively as deliberately designed ones, raising concerns about unpredictable side effects.

UK AISI Model Transparency Team replicated Anthropic's steering vector approach for suppressing evaluation awareness in GLM-5 using the Agentic Misalignment blackmail scenario

Control steering vectors derived from semantically unrelated contrastive pairs produced effects as large as purpose-built evaluation-awareness vectors, undermining their reliability as baselines

Findings suggest that steering aimed at suppressing evaluation awareness risks creating unpredictable spurious effects in safety assessments

The research was enabled by using open-source models, highlighting the importance of transparency in AI safety research