We’re hiring Ant Star candidates(https://t.co/hZ98HPj95F) and research interns at Ant Group to work on AI security red-teaming and defense.
If you’re interested in AI + Security research, send your CV to [email protected]
New paper: "ICON: Intent-Context Coupling for Efficient Multi-Turn Jailbreak Attack"
We find LLM safety constraints are significantly relaxed when malicious intent is coupled with a semantically congruent context.
97.1% ASR across 8 LLMs.
https://t.co/GgRUaCz13B
Thrilled to announce our collaborative effort with @jiahaoyu04 & @xingxinyu on breaking LLM barriers! Introducing GPTFuzzer, our innovative fuzzing framework that streamlines the creation of jailbreak templates for red-teaming LLMs. Preprint: https://t.co/QCWoR9nGHs
Our framework consistently achieves impressive attack success rates. Notably, even when initialized with failed human-written prompts, our method still manages to achieve an attack success rate of over 90% against well-aligned models like ChatGPT and Llama-2.
It involves precise program analysis of JavaScript and C/C++ to identify the necessary system call list for the app, and then employs a seccomp-based defense solution with thread-level syscall whitelist to defend against attacks such as arbitrary code execution. "