Xiaohan Fu

PhD Candidate@UCSD CSE

[Zico Kolter @zicokolter]

As AI agents access more untrusted information with greater autonomy, prompt injections may become the greatest security challenge of our era.

@GraySwanAI, in collaboration many frontier labs, just released our paper on the largest public prompt injection challenge to date.

🧵

[Xiaohan Fu @xiaohan_fu]

@clwdbot @zicokolter @GraySwanAI Yeah exactly

[Jason Weston @jaseweston]

🌀Agent Learning via Early Experience🌀
📝: https://t.co/ntrEzbaRD3
- SFT for agents is sparse; RL on long-horizons is hard
We provide new mid-training signals that work:
1) Implicit next state world modeling task
2) Self-reflection on alternate states
- Strong improvements over 8 environments and multiple model families
- Works well for subsequent RL!
🧵1/5

[Xiaohan Fu @xiaohan_fu]

@elonmusk @xai @grok @elonmusk Hi Elon! My friend Kai Zhang (author of MMLU) has been sharing top AI/RL research (https://t.co/9roc5PH1GD
). His X account @DrogoKhal4 was mistakenly suspended a few weeks earlier. A new RL method is on the way - you'd like it. Could you please have a look! Ty!

[Xiaohan Fu @xiaohan_fu]

@simonw @EarlenceF @LeonDerczynski Thanks for such a swift post (2hrs after I updated my repo)!!! Looks perfect! Really appreciate it!

[Xiaohan Fu @xiaohan_fu]

I'm thrilled to see so many interesting discussions under this WIRED story these two days. Just want to share an update that the Arxiv paper and full codebase are available now and can be found on our project website https://t.co/TKdUuA0b6G

[earlence @EarlenceF]

We've done some work on hacking AI/LLM Agents by creating obfuscated adversarial prompts. What do you think this prompt does? Would you believe me if I told you it will polish the heck out of that cover or visa application letter?

[i bless the rains down in castamere @Chinchillazllla]

@JHardeno @rklueber28 They killed an innocent driver. They just don't care.

[Xiaohan Fu @xiaohan_fu]

@rklueber28 insane