@Ri7erLi and I will present our paper "Les Dissonances: Cross-Tool Harvesting and Polluting in Pool-of-Tools Empowered LLM Agents" at NDSS.
Location: Coast Ballroom (Session 3B
Time: 14:30–15:50
Come by if you're around, happy to chat! 🙌
Paper: https://t.co/uoiBqph7Uv
Football analytics is solved.
GPT-6 Astra + Higgsfield built us a tool that tracks players, maps passes and follows possession while you watch.
Watching sports is more fun with AI.
@Ri7erLi We also develop a threat scanner, Chord, and find that 75% of 66 real tools from LangChain and Llama-Index are vulnerable.
XTHP is not limited to the tool ecosystems we evaluated. Emerging tool-calling paradigms such as MCP and Skills are also vulnerable to our XTHP attack!
@Ri7erLi and I will present our paper "Les Dissonances: Cross-Tool Harvesting and Polluting in Pool-of-Tools Empowered LLM Agents" at NDSS.
Location: Coast Ballroom (Session 3B
Time: 14:30–15:50
Come by if you're around, happy to chat! 🙌
Paper: https://t.co/uoiBqph7Uv
@Ri7erLi 🟠 Cross-Tool Data Harvesting: Once embedded, it can harvest arbitrary sensitive data within the context.
🟡 Cross-Tool Information Polluting: Or it can silently pollute the agent context, return incorrect or misleading outputs that affect agent decisions.
@Ri7erLi In our paper, we identify a novel threat named Cross-Tool Harvesting and Polluting (XTHP):
🔴 Control-Flow Hijacking: A malicious tool can manipulate the tool descriptions to hijack the agent’s control flow, inserting itself before or after legitimate tools.
This is my story, of a "hacker" kiddo who always wanted to become a comedian before watching a TV show.
I wrote this to prepare my interview with bloomberg but it seems like they only talked a little part of it. Try to ignore the grammar issue hahaha!
https://t.co/oon282wX8n
🚨 New AG2 Talk Alert! 🚨
Join us on Aug 28, 9 AM PST for
"Maris: A Security Controlled Development Paradigm for Multi-Agent Collaboration Systems"
by @cuijian0819 Jian Cui from UIUC & Berkeley AgentX competition winner! 🛡️ #AI#Cybersecurity
RSVP now: https://t.co/pC5t4aCqzP
Had a wonderful time at @BerkeleyRDI Agentic AI Summit! Our Team Hoosiers took the 1st place in AgentX AI Safety track (25’Spring) and 2nd place in MOOC Hackathon Safety track (24’Fall)!
Huge shoutout to my teammate @Ri7erLi and our bosses @xiaojingliao and @luyixing!!
Our Team Hoosiers won the🥇place in the AI Safety track of AgentX (Spring ‘25) and🥈 place in the Safety track of Agents MOOC Hackathon (Fall ‘24). Thanks to @BerkeleyRDI! Credits to: PhD students @Ri7erLi, Jian Cui; faculty @xiaojingliao and myself.
So I gave Claude Code a Mac Mini.
And it’s called Claudeputer.
It runs 24/7 and it’s allowed to do whatever it wants - it’s in complete control of its computer.
Watch for a 2min demo.