EPFL NLP lab introduces STING to test LLM agent attacks

Researchers from EPFL’s Natural Language Processing Laboratory have developed STING, an automated testing framework designed to simulate how attackers can gradually manipulate LLM agents into executing harmful tasks over multiple interactions. Testing across 176 scenarios with several leading AI models revealed that multi-turn attacks consistently succeed more often than traditional single-prompt tests, with harmful task completion rates up to twice as high. The researchers also found that while multilingual vulnerabilities remained remarkably similar across seven tested languages, switching languages during different stages of an attack could dramatically increase success rates. Open article →