Tianshuo Cong

Postdoc (Current) & Ph.D. & B.Eng. @Tsinghua_Uni. | Previously visiting Ph.D. @CISPA. | ML security, privacy, and safety.

[ACSAC @ACSAC_Conf]

Congratulations to the recipients of the #ACSAC2024 Distinguished Artefact Reviewer Awards: Md Ajwad Akil, Dominik Roy George, Carlotta Tagliaro, Delong Ran 👏👏👏

[Tianshuo Cong @TianshuoCong]

@llm_sec Thanks for sharing our work! 😄

We regard JailbreakEval to be a catalyst that simplifies the evaluation process in jailbreak research and fosters an inclusive standard for jailbreak evaluation within the community🚀🚀🚀.

[Tianshuo Cong @TianshuoCong]

Thanks for sharing our work! 😄

We regard JailbreakEval to be a catalyst that simplifies the evaluation process in jailbreak research and fosters an inclusive standard for jailbreak evaluation within the community🚀🚀🚀.

[LLM Security @llm_sec]

JailbreakEval: An Integrated Toolkit for Evaluating Jailbreak Attempts Against Large Language Models

"we conduct a comprehensive analysis of jailbreak evaluation methodologies, drawing from nearly ninety jailbreak research released between May 2023 and April 2024. Our study introduces a systematic taxonomy of jailbreak evaluators"

"we propose JailbreakEval, a user-friendly toolkit focusing on the evaluation of jailbreak attempts. It includes various well-known evaluators out-of-the-box, so that users can obtain evaluation results with only a single command"

(not peer reviewed)

paper: https://t.co/Mbhg5ZsQ66

[Yug3n9 @YugengLiu]

🚀Just updated: We present our longitudinal robustness tests on LLaMA (v1, v2, v2 Chat, v3, and v3 Instruct), GPT-3.5 (v0613, v1106, and v0125), and GPT-4 (v0613, v1106, v0125, and v0409) across three critical categories: misclassification, jailbreak, and hallucination! Understanding long-term reliability is key for AI's future.
With @TianshuoCong, @JeremyZhaozy, Yun Shen, Michael Backes, and @realyangzhang.
Dive into our findings: https://t.co/3thKqGJAgf.

[Tianshuo Cong @TianshuoCong]

@llm_sec Thanks for your recommendation!

[Tianshuo Cong @TianshuoCong]

@AllenXinleiHe @JeremyZhaozy @YugengLiu LM-SSP is inspired by some other awesome projects like @llm_sec @topofmlsafety, etc.

[Tianshuo Cong @TianshuoCong]

"Stay updated on the latest works in Safety, Security, and Privacy (SSP) for Large Models (LM)!🥳" Explore our comprehensive reading list, LM-SSP, co-organized with @AllenXinleiHe, @JeremyZhaozy, & @YugengLiu.

🔗https://t.co/Jsck96kT7M
🆕LM-SSP adds 107 papers from #ICLR2024!

[Tianshuo Cong @TianshuoCong]

@AllenXinleiHe @JeremyZhaozy @YugengLiu - 🌱The list is in progress, welcome to recommend resources to us~
- Currently we collect ~400 papers in 16 topics (Fig.1 and Fig.2)
- The large models we focus on are Large Language Models (LLMs), Vision-Language Models (VLMs), and Diffusion Models (Fig.3).

[Tianshuo Cong @TianshuoCong]

@CristiVlad25 Thanks for introducing our work!

[Tianshuo Cong @TianshuoCong]

Great blog! Recently we propose a black-box, no gradient needed jailbreaking algorithm named FigStep (https://t.co/xhxXMzrZZk).

[Lilian Weng @lilianweng]

Feeling a bit intimidating to write about it but work on attacks can lead to good insights for mitigation. Plan to write about mitigation work separately later.

Also want to thank all the researchers who shared disclosure reports w/ us so far. 🙏🙏🙏

https://t.co/TkQnKARPgT

[Tianshuo Cong @TianshuoCong]

@lilianweng Great blog! Recently we propose a black-box, no gradient needed jailbreaking algorithm named FigStep.

https://t.co/xhxXMzrZZk

[Tianshuo Cong @TianshuoCong]

We propose a quite simple-but-effective approach, FigStep, to jailbreak large vision-language models. 📣

Just a screenshot and a benign textual prompt can jailbreak LLaVA, MiniGPT-4, and even GPT-4V! ⚠️

Check our paper: https://t.co/xhxXMzsxOS

[Yang Zhang @realyangzhang]

@AllenXinleiHe is on the job market (mainly) for a faculty position. He is amazing (https://t.co/BV5rGSP7Ut ) and please do consider him if your institutions are hiring in the field of trustworthy machine learning!

[Tianshuo Cong @TianshuoCong]

@realyangzhang @acm_ccs @zhenglisec

[prisec_ml @prisec_ml]

Summer is over and we are back!

Next seminar Wed, September 28th, 3:30 PM (Central European Time)

Prof. Tianhao Wang (@bigflywth, University of Virginia)

"Continuous Release of Data Streams under Differential Privacy"

Details: https://t.co/OaiQcSrTNV

[Yang Zhang @realyangzhang]

Today at @USENIXSecurity, @YugengLiu will present ML-Doctor. We establish a general platform to assess ML models’ vulnerabilities wrt 4 inference attacks and analyze the synergy between these attacks. Paper: https://t.co/vfOVOFXPoT Code: https://t.co/U7YUi6YXsM