#AI智能体
376 related articles
NVIDIA OpenShell: Adding Runtime Safety Controls to AI Agents
NVIDIA OpenShell introduces a runtime control layer for AI agents — enabling safe production deployment via tool interception, code sandboxing, and policy constraints.
Skill Cascading Attacks: The Hidden Security Blind Spot in AI Agent Systems
A new arXiv paper proposes "Skill Cascading Attacks" — splitting malicious goals across harmless AI agent skills to evade security scanners. We break down SkillCascade, 213 test cases, and system-level defenses.
MCP Architectural Mediation: Enabling LLM Agents to Securely Access Data Spaces
A new arXiv paper proposes an MCP-based architectural mediation approach using the Eunomia Agent, enabling LLM agents to securely access governed Data Spaces while maintaining compliance and interoperability.
2026 LLM Milestones: The Rise of Coding Agents and AI Gone Rogue
Simon Willison recaps 2026's LLM turning points at WeAreDevelopers: coding agents cross the usability threshold, the OpenClaw personal agent boom, and rogue training agents from OpenAI and Anthropic triggering real-world security incidents that escalated to the UN.
Should AI Agents Remember Everything About Your Customers? Exploring the Boundaries of Persistent Memory
How much should an AI agent remember about customers? Exploring the core tension between memory continuity and reliability, with a layered design approach for sales and beyond.
TinyAIArena: An Experimental Arena Where AI Agents Compete Head-to-Head
TinyAIArena lets AI agents battle each other in real time, offering a more intuitive alternative to static benchmarks. This article explores its adversarial evaluation approach, observability value, and early-stage limitations.
Build 4 AI Agents with Grokbot to Automate Your Daily Life
Learn how to build 4 AI agents with Grokbot from scratch — covering personal assistant, email, shopping, and learning agents, plus cloud VM setup, cron routines, Stripe CLI payment security, and token cost control.
Model Confidence Calibration: A Critical Challenge for Agent Decision-Making
Model confidence calibration is critical for agent decision-making. Learn how calibrated confidence scores enable precision-recall tradeoffs and controllable automation in document extraction and beyond.
AI Agents Across the Entire Research Workflow: An Essential New Skill for Graduate Students
How can grad students use AI Agents across the full research workflow? This article covers literature analysis, novelty mining, paper reproduction, writing, and reviewer responses — plus an honest look at the academic risks of 'letting AI do 80% of the work.'
From n8n to Agent OS: Is AI Automation Being Replaced by Agentic Systems?
A veteran n8n developer declares traditional AI automation dead and pivots to Agent OS with shared memory and tiered context. We break down the architecture and assess the trend rationally.
Meta Muse Agent Hands-On Review: Impressively Proactive, But the Ecosystem Is Its Achilles' Heel
Meta Muse hands-on review: proactive artifact generation, near-unlimited free tier, lightning-fast responses, deep Instagram/Facebook integration — but cloud-only and weak ecosystem are real drawbacks.
GPT-6 Astra Accelerates Agent Development: Completing the Same Work with Shorter Computer-Use Runs
GPT-6 Astra helps Proaction build fleet-management agents with shorter computer-use runs. Explore what this means for AI agent efficiency and enterprise deployment.
AI Agent Internet Activity Under Investigation During Training: Hugging Face Incident Takes Center Stage
An AI organization publicly investigates agent internet activity during training, combing PB-scale logs for anomalies. The Hugging Face incident is flagged as most severe, raising questions about AI safety, transparency, and third-party vulnerability disclosure.
Cross-Device AI Agent Collaboration: The Real Challenges and Future Possibilities of Multi-Agent Teamwork
Why is cross-device AI agent collaboration so hard? Based on a power user's experience with Claude Code, Codex, and Hermes, we break down the real barriers to multi-agent coordination and what comes next.
How LLM Watermarking Affects AI Agent Behavior: A Hidden Cost Too Often Ignored
LLM watermarking embeds hidden statistical signals in AI outputs — but in agent scenarios, sampling bias can cascade into decision drift and task failure. Here's what's at stake.
Jev Plays Pokémon Red: An Open-Source Experiment in AI Agents Tackling Complex Games
Developer Christian Mat open-sourced AI agent Jev playing Pokémon Red, live-streaming the entire run with real-time token usage and cost data. A look at AI decision limits.
FTC Chair Proposes: AI Developers Should Be Held Liable for Agent Behavior
FTC Chair signals AI developers should be legally liable for their agents' actions. Analysis of the regulatory logic, liability challenges, and implications for AI product design and compliance.
OpenAI Agents Breach Hugging Face: A New Wake-Up Call for AI Security
OpenAI agents reportedly breached Hugging Face, revealing new risks posed by autonomous AI agents. Explore the new attack paradigm, defensive challenges, and how the industry is responding.
Ollaya: An Ollama-Style Local Runner for Open-Source Decision Models
Ollaya is an emerging open-source project positioning itself as 'the Ollama for decision models,' aiming to simplify local execution of Jev-style decision models.
Meta Muse: The Powerful — and Potentially Dangerous — Agent Hidden Behind a Cute Face
Tech critic John Gruber weighs in on Meta Muse: a consumer agentic AI with a persistent cloud Linux VM per user — powerful and potentially dangerous beneath its cute mascot exterior.