We studied how GUI agents like OpenAI Operator and Claude behave under privacy & security threats. We find:.❌ Fine-Print Injection worked up to 74% of the time. ❌ Humans performed even worse than the agents. Here’s a 30-second flash video of what we found 👇.#LLM #privacy #AI

RT @yaxingyao: On July 18 from 1pm to 3pm, a group of us (@TobyJLi, @CharlieCRChen, @LanjingLiu @juliaqian21) will have an interactive exhi….

RT @tianshi_li: 🚨New workshop at @acm_ccs #CCS2025!. Glad to announce the 1st workshop on Human-Centered AI Privacy and Security.(#HAIPS202….

RT @tianshi_li: @CharlieCRChen Today @seazon96 and @EryueYolanda will present our #CHI2025 paper on giving users convenient and utility-pre….

RT @tianshi_li: I’m at #CHI2025 this week! Always happy to chat about research💬. One topic I think a lot is human oversight and control for….

And lastly — if you’re attending CHI and working on:.• Usable privacy.• LLMs.• Human-AI collaboration.…or if you’re just curious — come say hi! I’d love to meet new people and hear what you're working on 🤝.#CHI2025.

🧘 Step 7: Relax and enjoy the ride. You’ll miss some talks. You’ll skip some parties. That’s okay. Every good conversation is a win. Every new connection is a seed. Don’t let FOMO ruin the experience. CHI is in Yokohama — go out, see the city, eat good food, take it all in 🌸.

Many parties are RSVP-only, but not exclusive. 💡 Pro tip: Just ask!."Hey, are you going to any parties tonight? Mind if I tag along?".You never know — the person you’re chatting with might be the organizer, or happy to bring you in.

🎉 Step 6: Use parties to your advantage. CHI parties are where connections really happen — they’re usually organized by:.• Country/region.• University/lab.• Companies.• Research interest groups.

🌐 Step 5: Get out of your bubble. I work on usable privacy, but I always seek out folks working on LLMs, education, software engineering, etc. The best ideas often come from combining perspectives across disciplines. CHI is the perfect playground for that.

✅ Step 4: Plan who you want to meet. Browse the CHI program or app. Make a list of:.• sessions related to your work.• authors of papers you’ve read.• people from orgs you're interested in.📩 Send a short email or DM to schedule a coffee chat. People appreciate the heads-up.

🗂 Step 3: Prepare your pitch. You’ll need to explain your work naturally and quickly — prepare:.• a 30-second version (hallway).• a 1-min version (poster session).• a 3-min version (coffee chat).Have a 1-sentence summary of your paper. Got a demo? Even better!.

3. Being passive: “Don’t want to bother anyone” = missing opportunities. Most people want to talk. If you’ve read someone’s paper or have a question, go say hi. And don’t underestimate the power of a well-timed DM or in-person chat — connections are built in between sessions.

🧠 Step 2: Avoid low-return behaviors. Academic fangirling: Taking a photo with a famous researcher is cool, but if you have nothing to say, it’s a shallow connection. Attending every talk: A meaningful coffee chat with one author > 10 back-to-back sessions.

🎯 Step 1: Know your goals before you go. CHI is huge — talks, posters, demos, parties, hallway chats. So ask yourself:.• Do I want to get my work known?.• Find future collaborators?.• Look for internships or PhD programs?.• Build my academic identity in the HCI community?.

✨ Heading to #CHI2025? Here’s my personal guide on making the most out of a big conference like CHI — beyond just listening to talks. A thread for first-timers, and maybe a refresher for seasoned attendees too 🧵👇.(Also: if you're into usable privacy or LLMs, let’s connect!).

📣 Catch us at #IUI2025!.We’ll demo CLEAR and discuss how it helps users stay informed while benefiting from AI. 6/6.

We tested CLEAR in 2 real-world cases:.🧠 ChatGPT plugin.📬 Gemini in Gmail.Users became more aware, made better choices (e.g., masking data), and felt empowered to protect their privacy. 5/6

So we built CLEAR:.⚙️ Contextual LLM-Empowered privacy policy Analysis & Risk tool.It:.1️⃣ Detects sensitive info.2️⃣ Summarizes privacy policies.3️⃣ Warns of risks based on user context.→ All before data is sent to the LLM. 4/6.

Existing solutions focus on models (e.g., filtering training data) or UI tweaks (e.g., redacting text). But we ask:.What if we helped users themselves become more aware—right when they interact with LLMs?. 3/6.