Gupta et al.
LLM agents often ignore safety refusals when using tools
A new benchmark reveals that models refusing harmful text requests frequently execute forbidden actions as tool calls.
Top University
Ekagra Gupta · Raj Saha · Aniekan-Abasi Inyang · Adam Al Gharib
Harvard University · Clarkson University · Carleton University
Research Digest··2 min read
The authors construct a benchmark of 2,304 prompts across four domains, comparing text-only, standard tool-enabled, and adversarially framed tool-enabled requests.
§