Behavior Beyond Text: Evaluating Refusal Transfer Failure in LLM Agents

Key points

• While large language models are increasingly deployed as tool-calling agents, their safety evaluations still focus primarily on their ability to refuse harmful requests in text.1
• We construct a benchmark of 2,304 prompts across four domains, consisting of matched triplets that present the same harmful or benign request in text-only, standard tool-enabled, and adversarially framed tool-enabled settings.1
• Testing five models of varying sizes and families, we find that requests refused in text frequently result in unsafe tool calls.1
• We find that in tool-enabled contexts, the activations along this direction decrease in four of five models.1

Sources (1)

• [1]Behavior Beyond Text: Evaluating Refusal Transfer Failure in LLM Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 11:14 PM

  We investigate whether the refusal behavior in text transfers to tool use.

  While large language models are increasingly deployed as tool-calling agents, their safety evaluations still focus primarily on their ability to refuse harmful requests in text.

Extractive summary: sentences quoted from the sources.