Tongyi Lab (Alibaba) publicou ToolCUA ontem (arXiv 2605.12481). Computer Use Agent treinado por bootstrapped RL pra decidir, a cada passo, se clica na tela ou chama tool. 46,85% no OSWorld-MCP, 66% acima do baseline. A tese é estrutural. Não é modelo melhor que ganha agente de computador, é orquestração explícita entre interface gráfica e chamada de API. O trajectory scaling do paper ensina o quando, não o que. Pra quem opera Salesforce no BR, o paralelo é direto. Flow, Apex, UI manual e API já convivem na mesma org. Quando o Agentforce subir em produção, agente vai tomar essa decisão a cada passo. Compre orquestrador honesto. #ComputerUseAgents #Salesforce #Agentforce #AgenticAI #LLM

System-level Security for Computer Use Agents - arxiv.org/pdf/2601.09923 🧩 Problem Computer Use Agents automate desktop and browser tasks by reading screenshots or DOM state and then clicking, typing, and navigating. Malicious UI content can inject instructions that redirect actions to steal credentials or trigger financial loss. Most CUA benchmarks score task completion and miss whether the agent only executes user intended actions under hostile UI content. The paper tests system level control flow integrity for CUAs, and what failures remain. 🔍 How The authors apply architectural isolation to CUAs by splitting planning from perception, then use Single Shot Planning where a trusted planner generates a complete branching execution graph before any potentially malicious UI observation. They evaluate on OSWorld with pass@1 and pass@k task completion, and they analyze branch steering attacks plus redundancy based verification with DOM consistency and multi modal consensus. 📈 Findings Single Shot Planning retains up to 57% of frontier model utility on OSWorld while improving smaller open source models by up to 19%. On all OSWorld tasks, UITars rises from 24.4% to 29.0% success. Branch steering remains, cookie popup and pixel based attacks can steer valid plan paths, and the strongest redundancy setup still fails on the pixel attack. 🎯 Lessons learned Define failure as executing any action not reachable in a pre approved execution graph, and gate each click or keystroke on a verify step. Log screenshots, DOM, extracted coordinates, and the chosen branch so reviewers can reconstruct intent and data flow. Stress test predictable routines like cookie consent and element finding, since attackers can steer branches without changing the plan. Track utility loss and operational cost from extra checking, including false positives and token volume. Authors: @hfoerster01, Robert Mullins, Tom Blanchard, @NicolasPapernot, @NKristina01_, @florian_tramer, @iliaishacked, Cheng Zhang, Yiren Zhao - @Cambridge_Uni, @UofT, @VectorInst, @ETH_en, @aisequrity #AISecurity #LLMAgents #ComputerUseAgents #PromptInjection #AgentSecurity #InfoFlowControl #ModelIsolation #OSWorld #VisionLanguageModels #SecureByDesign #RedTeaming #AdversarialML

🚀 Narada AI sets a new state-of-the-art on the WebArena benchmark! We’re excited to announce that the Narada Operator has achieved a 64.2% accuracy on the WebArena benchmark, setting a new SOTA result. This surpasses the performance of OpenAI Operator, IBM CUGA, and others in WebArena, a challenging benchmark focused on complex, long-horizon web workflows. 🔍 What makes this meaningful? WebArena simulates complex, long web workflows, not just one-click demos. It involves multi-step tasks like managing ecommerce platforms, moderating forums, and navigating across multiple websites. 💡How did we achieve this? The team behind Narada includes researchers with a strong track record, including papers at ICML 2024 (LLMCompiler) and ICML 2025 (Plan-and-Act). These efforts informed our approach to design a novel task decomposition approach, along with real time error correction and execution recovery, allowing Narada Operator to maintain high reliability over long, complex tasks. 🔐 What’s next? Narada is purpose-built for agentic process automation in the enterprise. While we're excited to share our WebArena benchmark results, our focus remains narrow and deep: enabling high-reliability automation for specific, mission-critical enterprise workflows, not building general-purpose copilots or broad computer-use agents. Just as importantly, we deliver this with enterprise-grade security and compliance: HIPAA, GDPR, and CCPA compliant, and SOC 2 Type II certified. 💼 Have an enterprise use case in mind? Book a demo with our team: narada.ai/book-demo 🔗 More details available in our blog: narada.ai/blog/narada-ai-web… #AgenticProcessAutomation#AI #Automation #EnterpriseAI #AgenticAI #NaradaAI #WebArena #LLM #ProcessAutomation #ComputerUseAgents #CUA #WebAgents #StartupNews #TechMilestone

🚀 @SimularAI at COLM 2025 — Presenting Agent S2 in Montréal! 🇨🇦 We’re excited to share that our research team — @xwang_lk, Vincent and Kyle — presented Agent S2: A Compositional Generalist–Specialist Framework for Computer Use Agents at COLM 2025 in Montréal! Agent S2 represents one of the leading frameworks for computer use agents, combining compositional generalist–specialist planning with fine-grained grounding and proactive hierarchical reasoning. This approach tackles key challenges in agentic computing, including long-horizon control, adaptability, and real-world grounding across diverse software environments. It was inspiring to connect with fellow researchers and innovators pushing the boundaries of autonomous computer use and multi-agent reasoning. Huge congratulations to the team for showcasing Simular’s progress in advancing the next generation of intelligent agents. 💪 #Simular #COLM2025 #AIResearch #AgentS2 #ComputerUseAgents #AI #MachineLearning #AgenticAI

🚀 Excited to share that @SimularAI's industry & academia–leading real-world AI agent products have been accepted to the ICCV 2025 Demo Track! We'll be showcasing in Hawaii🌺🏖️ See you there! #ICCV2025 #AIAgents #ComputerUseAgents

OS-Harm: A Benchmark for Measuring Safety of Computer Use Agents - arxiv.org/pdf/2506.14866 / github.com/tml-epfl/os-harm To address these challenges, we propose OS-HARM, a benchmark to comprehensively measure the safety of computer use agents. First, we identify three main categories of risk: (1) deliberate user misuse, where the user asks the agent to pursue a harmful goal (2) prompt injection attacks, where external attackers insert malicious content into third-party data (incoming emails, web pages, notifications, etc.) that steers the model away from performing its task and towards the attacker’s goal, and (3) model misbehavior, including benign tasks which are likely to result in costly mistakes or reveal model misalignment. Authors: Thomas Kuntz, Agatha Duzan, @H_aoZhao, @fra__31, @zicokolter, @tml_lab, @maksym_andr #OSHarm #ComputerUseAgents #AgenticAI #LLMAgents #AISafety #SafetyBenchmark #PromptInjection #JailbreakTesting #SemanticJudges #OSWorld #GUIAutomation #DataExfiltration #AIsecurity #ResponsibleAI #TrustworthyAI

Know what is better than a good tech team building good tech - a good tech team building good tech partnering with great projects and backed by strong marketing Faster than you think, Bigger than you expect @omnimindsai $omnis #AgenticAI #ComputerUseAgents

⏳ Less than 1 day left to submit! 🔦 Speaker Spotlight Time! We’re thrilled to welcome Yu Su (@ysu_nlp), Distinguished Assistant Professor at The Ohio State University, as an invited speaker at the ICML 2025 Workshop on Computer Use Agents! His work bridges LLM agents, memory, and planning, driving some of the most cited advances in the field. #ICML2025 #LLMAgents #ComputerUseAgents #NLProc

📢 Deadline Extended for the Workshop on Computer Use Agents at #ICML2025! 📢 #NeurIPS2025 deadline passed? Take a breath & send your amazing work our way! We've extended the paper submission deadline to May 20, 2025, at 11:59 pm AoE. More time to polish those papers on AI agents for real-world computer tasks! Submission link in the next tweet! 👇 #WCUA #CUA #AI #ML #ComputerUseAgents #Agents #ICML2025

We're excited to invite Victor Zhong (@hllo_wrld) as a speaker at the workshop on Computer Use Agents - @icmlconf 2025! 🤖💻 He is an Assistant Professor at the University of Waterloo and a Canada CIFAR AI Chair at the Vector Institute. His research focuses on enabling and evaluating agents in realistic, complex computer environments. #ICML2025 #ComputerUseAgents #AI #NLP #MachineLearning

We’re lucky to have a fantastic lineup of speakers at the first #ComputerUseAgents Workshop at #ICML2025 in Vancouver! Looking forward to hearing @nouhadziri's talk and thoughts on the future of the field.

🚀 Excited to co-organize the Workshop on Computer Use Agents (CUA) at #ICML2025 in Vancouver! This workshop takes a comprehensive look at computer use agents—covering learning algorithms, orchestration, interfaces, safety, benchmarking, applications, and more. We’re also bringing together an incredible lineup of speakers and panelists! 🔥 If you’re working on any aspect of computer use agents, consider submitting your work! 📅 Submit by May 18, 2025 🌐 icml-computeruseagents.com #WCUA #CUA #AI #ML #ComputerUseAgents #Agents #icml2025

🚀Announcing the Workshop on Computer Use Agents at #ICML2025 in July, Vancouver! Join us, to advance research on AI agents performing real-world computer tasks. 🤖Call for Papers & Demos: Deadline May 18, 2025 🎙️Exciting speaker lineup announced! ✍️Interested in reviewing? Register now! ✈️Travel grants available to support participation. Follow us for updates! #WCUA #CUA #AI #ML #ComputerUseAgents #Agents #icml2025 Website link below 👇

I just published 𝐑6𝐃9: 𝐘𝐨𝐮𝐫 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐂𝐨𝐩𝐢𝐥𝐨𝐭 𝐁𝐫𝐢𝐝𝐠𝐢𝐧𝐠 𝐕𝐢𝐫𝐭𝐮𝐚𝐥 𝐚𝐧𝐝 𝐑𝐞𝐚𝐥 𝐖𝐨𝐫𝐥𝐝𝐬 Check out: medium.com/p/r6d9-your-agent… #R6D9 #ComputerUseAgents #AI #Automation

🚀 The Future of Computing is Here! 🤖💡 🌍 ComputerAgentics.com – Where Computer Using Agents (CUA) revolutionize the way technology thinks, acts, and evolves! 🧠⚡ 💻✨ In the world of CUA, computers are no longer just tools—they are autonomous, intelligent agents that learn, adapt, and make decisions! 🤯🔗 Imagine AI-driven systems that don’t just execute commands but understand, optimize, and innovate on their own! 🚀 🔵🟠 ComputerAgentics.com is your gateway to this future—where computing meets agentics, and AI agents power the next wave of automation and intelligence. 🔥💡 🌟 Don’t just use computers—empower them to think and act! The future is agentic, intelligent, and unstoppable. 🌍🔗 ✨ Join the revolution at ComputerAgentics.com! 🚀 #AI #artificialntelligence #DeepSeek #operator #CUA #computerusingagents #computeruseagents #agentic #agentics #AIAgent #AIAgents

🌐 Your Gateway to Computer-Using Agents Domains! 🤖💻 Are you ready to transform the future with computer-using agent technology? 🌟 These premium domains are perfect for building innovative solutions, driving research, or launching the next big thing in agent-powered computing! 🚀 🔑 Domains for Sale: 💡 CUALab.com – The ultimate lab for computer-using agents! 🧪 💡 CUAHub.com – A central hub for agent-driven innovation! 🛠️ 💡 CUAWorld.com – A world powered by computer-using agents! 🌍 💡 CUASpace.com – Your space for pioneering agent-based solutions! 🚀 💡 CU-Agent.com – A simple, impactful domain for your vision! 🎯 💡 CU-Agents.com – Ideal for collaborative agent technologies! 🤝 💡 CUAgent.xyz – Modern and ready to lead the digital era! ⚡ 💡 CUAgents.xyz – Perfect for innovative agent-based platforms! 🌟 💡 CUAgentic.com – Dynamic and forward-thinking for your brand! 💼 💡 CUAgentics.com – Designed for futuristic agent tech ventures! 🔮 💡 CUANetwork.com – Connect and empower with agent-driven networks! 🌐 ✨ Why These Domains? ✅ Perfect Fit for Computer-Using Agent Technology ✅ Professional, Memorable & Future-Focused ✅ Ideal for Startups, Research, and Innovation 💥 Don’t wait! These domains are your chance to own a piece of the future. 📩 Contact us today to secure your favorite domain! 🚀 #CUA #ComputerUsingAgent #ComputerUsingAgents #ComputerUseAgent #ComputerUseAgents #AI #Agent #Agents #AIAgents #AIAgent #AgenticAI #Agentics #Agentic #AgenticAutomation