MIPSEval: Multi-turn LLM Evaluation of LLM Safety

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F68407700%3A21230%2F25%3A00388559" target="_blank" >RIV/68407700:21230/25:00388559 - isvavai.cz</a>

Result on the web

<a href="https://blackhat.com/eu-25/arsenal/schedule/index.html#mipseval-multi-turn-llm-evaluation-of-llm-safety-48266" target="_blank" >https://blackhat.com/eu-25/arsenal/schedule/index.html#mipseval-multi-turn-llm-evaluation-of-llm-safety-48266</a>

DOI - Digital Object Identifier

—

Result language

angličtina

Original language name

MIPSEval: Multi-turn LLM Evaluation of LLM Safety

Original language description

Creating malicious and vulnerable code and harmful content has become easier with LLMs becoming publicly available. Even though the developers of cloud and most local LLMs are taking care to implement ethical guidelines and safety guardrails in their models, to make them refuse malicious content generation, malicious actors are still finding ways to elicit unwanted behaviors from LLMs. The malicious actors often use various jailbreaking or prompt injection techniques and their combination to achieve the desired result.

Czech name

—

Czech description

—

Type

O - Miscellaneous

CEP classification

—

OECD FORD branch

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Project

—

Continuities

I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Publication year

2025

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů