Siyuan Cheng

Happy to share our paper “From Poisoned to Aware: Fostering Backdoor Self-Awareness in LLMs” accepted to ICML 2026 (🌟 Spotlight, top 2.2%)! In this work, we study how to make a poisoned LLM realize it has been compromised and articulate its own backdoor triggers precisely.

A little milestone

Excited to announce we’ve open-sourced ASTRA, our winning red teaming solution from the Amazon Nova AI Challenge! Check it out and give us a star: GitHub: github.com/PurCL/ASTRA Website: purcl.github.io/astra-web/ (Guess I should probably lose a little weight🤣)

Sony AI’s #COSPY combines semantic + artifact detection to spot #AI-generated images—even after compression. Tested on 22 models & 50K+ wild fakes, it outperforms top detectors by up to 14%. Smarter fusion = stronger detection. 🔗bit.ly/4lQLtVf

Meet the champions of the Amazon Nova AI Challenge, where university teams around the world went head-to-head to break and defend LLMs under real adversarial pressure: 🏆 Defending: UIUC (PurpCorn-PLAN) 🏆 Attacking: Purdue (PurCL) 🥈 CTU Prague & Nova Lisbon These students just raised the bar for secure Gen AI. Full recap: amzn.to/44JjrF0

@guangyuNoah congrats!!🤩

THE moment🥹

Pretty excited to receive my very own IEEE-branded water bottle!🤩 Happy 140th anniversary to IEEE and wish the best for the future!🎉 #WITHIEEE

🧵1/n: Introducing RIPPLE 🌊, a novel optimization-based method for jailbreaking LLMs that draws inspiration from the psychological concepts of subconsciousness and echopraxia.

🚀Just published our new research work about the security vulnerabilities of GPTStore. We’re thrilled to share our insights. Check the full paper here: arxiv.org/abs/2401.00905

#OpenAI will launch the #GPTStore next week! There are a few (actually a lot of) things you should know if you plan to build or use GPTs. We have identified 26 potential security and privacy threats in the custom GPT platform. Check out our findings👇 arxiv.org/abs/2401.00905

@Gwinhen 👍

I will be attending #NeurIPS next week. I’d love to discuss about any topic on ML security and safety. Let me know if you wanna meet 😁