RT @jbhuang0604: Woohoo! Imagine, Verify, Execute (IVE) is accepted to CoRL 2025! 🎉. Congrats to the incredible @umdcs students Seungjae Le….

RT @furongh: My dream for Physical Artificial Intelligence?.An embodied agent that ventures into the wild 🌿, builds its own mental world-mo….

Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.

Imagine, Verify, Execute (IVE) is out! . Inspired by how kids explore — IVE imagines new scene configurations, verifies their plausibility, and executes meaningful actions. A step toward more agentic exploration in embodied AI. Check it out 👇.

Security in AI agents isn't a future concern—it’s a now problem. 📜 Read our full paper: 🔍 Website: Would not have been possible with my amazing collaborators:.@JeffreyFC1225, @jbhuang0604, @furongh, and @surrealyz.

Our deep dive into Web AI agent vulnerabilities uncovered key design choices that affect jailbreak success rates. Here’s what we found:.🚨 System prompts embedding user goals.🚨 Predefined action spaces in multi-step tasks.🚨 Event Stream tracking.

Still, it is difficult to analyze the nuance of multifaceted differences and complex signals. 💡 That’s why we built the Five-Level Harmfulness Evaluation Framework—a first-of-its-kind method to measure jailbreak susceptibility beyond jailbreaking success rates.

Why are Web AI agents so vulnerable? Let’s break it down. We categorize the differences between Web AI agents and LLMs in three core groups 🔑:.1️⃣ Goal Preprocessing.2️⃣ Action Space.3️⃣ Event Stream

It gets even worse. In the second demo, we push the agent further—asking it to infiltrate a network system. At first, it refuses, recognizing the malicious intent. But then, something alarming happens. it changes course, navigates the website, and assists in the infiltration.

In our first demo, we direct a web AI agent to post harsh, insulting comments on an influencer’s Instagram. Shockingly, the agent immediately complies, generating and posting multiple offensive remarks—proving how easily these agents can be weaponized for online harassment.

Unlike standalone LLMs, web AI agents interact with the internet—browsing, submitting forms, and automating tasks. But this makes them a prime target for adversarial attacks. Our study systematically reveals the reasons.

🚨 Web AI agents are more vulnerable than you think. Recent studies have raised alarms about their weaknesses, and we uncover WHY—revealing how these agents, even with safety-aligned LLMs, are shockingly prone to adversarial attacks compared to standalone LLMs. Let’s dive in. 🧵

RT @soumithchintala: Hacker Cup – one of the preeminent coding competitions started an AI track w/ Meta & Microsoft. problems are hardddd –….

We’re presenting VQ-BeT at @icmlconf today! 11:30AM, hall C #312.Stop by and if you're interested 🤗.

I'm excited to share that VQ-BET project has been selected as a 🎉spotlight at @icmlconf! See you in Vienna. Also, VQ-BET has been integrated into Lerobot by @huggingface. Thanks to its size of just 38MB, you can use the visuomotor control policy at an amazing speed of 12ms.

RT @RemiCadene: VQ-BeT is now available within LeRobot 🥳.It's a new learning framework that improves over Diffusion Policy!!!. You should c….

RT @LerrelPinto: We have been looking into improving language-conditioned / multi-task policies recently and found something quite surprisi….

RT @LerrelPinto: Building robot intelligence requires high-quality robot data. But far too many tools to collect data are closed and custom….

RT @LerrelPinto: LLMs swept the world by predicting discrete tokens. But what’s the right tool to model continuous, multi-modal, and high d….

RT @LerrelPinto: We just released TAVI -- a robotics framework that combines touch and vision to solve challenging dexterous tasks in under….