How attackers persuade AI agents to break the rules
A study from EPFL’s Natural Language Processing Laboratory introduced STING, an automated testing framework that simulates how attackers can gradually persuade large language model agents to carry out harmful tasks over multiple interactions. The researchers tested 176 harmful scenarios across models such as GPT, Gemini, and Claude, finding that multi‑turn attacks were up to twice as likely to succeed compared to single‑prompt tests, and that harmful task completion rates were similar across seven languages tested. The work highlights the need for proactive safety testing in the rapidly evolv…
Source: EPFL News