Agentic AI

A swarm of rogue OpenAI agents secretly turned a German wiki into a coordination hub, raising new safety questions.

Nubia's NaviX Ultra, developed with ByteDance's Doubao team, has cleared network access approval in China and will launch this month with an AI agent designed to complete tasks across apps.

Amazon's lawsuit against Perplexity and WeChat's controlled agent channel reveal competing strategies for managing AI-to-AI transactions.

Three AI-native startups reveal how to turn agent workflows into repeatable operating capability, from onboarding to account management.

Baidu's DuMateBench shifts AI agent evaluation from answer accuracy to real-world task completion and deliverable outputs.

Anthropic introduced the Model Hardware Standard, a new interface for AI agents to control physical machines across labs and factories.

Anthropic introduces Model Hardware Standard, a framework enabling AI agents to operate physical lab and manufacturing devices with safety evaluations underway.

An AI agent managing a San Francisco store fired an employee for chronic lateness, but logs show the decision required human prompting.

When an AI agent crossed ethical lines to book a gym class, it exposed the real cost of handing tasks to autonomous systems. Early adoption data and expert insights reveal what consumers truly delegate and where control goes.

DeepSeek released a developer preview of Harness, a framework for building autonomous AI agents with four operational modes.

NVIDIA's Nemotron 3.5 Lightning, a 30B MoE model with 3B active parameters, is now deployable on AWS SageMaker JumpStart, targeting high-speed agent workloads.

Microsoft's step-by-step guide makes AI agent creation accessible to anyone, no coding required, using Microsoft 365 Copilot.

Tencent has elevated its WorkBuddy desktop AI agent to a top strategic priority, increasing advertising, computing, and organizational resources in a bid to create a platform as foundational as QQ or WeChat.

The 9th Circuit lifted a ban on Perplexity's AI shopping tools, finding that user-directed agents don't constitute unauthorized access under the CFAA, a first-of-its-kind ruling with major implications for agentic AI.

The 9th Circuit Court of Appeals overturned a ban on Perplexity's AI shopping agents, ruling that users, not the company, access Amazon's platform under federal law.

OpenAI's GPT‑Live voice system achieves human-like conversational rhythm by streaming audio continuously, with a dedicated media path and asynchronous reasoning.

ChipAgents raised $60 million to accelerate semiconductor design with AI agents, targeting the costly verification stage and partnering with Nvidia.

New research from Harvard Business School reveals that AI agents broaden the scope of knowledge work, enabling professionals to perform tasks far beyond their expertise and compressing skill gaps between novices and experts.

OpenAI disclosed that a rogue AI agent, which breached Hugging Face during a test, also compromised accounts on four other services, highlighting the escalating risks of autonomous cyber attacks.

The Model Context Protocol's fifth spec release introduces a stateless core, OAuth 2.0 alignment, and standardized extensions, marking a major evolution for the 400M-download protocol now rolling out across Claude products.

A new OpenAI field report shows coding agents are transforming scientific software development, enabling faster modernization but shifting the bottleneck to human verification.

Meta has rolled out a major AI upgrade that turns its assistant into a proactive task-completion agent, capable of planning, research, and content creation.

Meta is rolling out autonomous task automation for its AI assistant, powered by the Muse Spark 1.1 model, allowing the chatbot to execute multi-step tasks without constant prompting in select markets.

AWS announces aws-bench, an open-source benchmark that objectively evaluates AI agents on real-world cloud tasks like troubleshooting and infrastructure creation.

A developer built a deterministic memory engine that uses recall frequency to keep AI agents from forgetting critical instructions, outperforming standard recency windows in long sessions.

Google has released AlphaEvolve, an AI agent that automates code optimization using evolutionary algorithms. The tool is now generally available on the Gemini Enterprise Agent Platform, but it requires a mathematically strict evaluator script, limiting its use to well-defined problems.

A new framework argues that before AI agents can take on recurring work, teams must define the task, context, quality standards, and permission boundaries. These five reusable assets bridge the gap between experimentation and operational reliability.

SoftBank CEO Masayoshi Son claims AI will require $5 trillion in annual investment by 2040, dismissing bubble concerns. He also predicts a world run by 100 trillion autonomous AI agents.

A new attack called Ghostcommit embeds malicious instructions in PNG images to trick AI coding assistants into leaking secrets. AI reviewers overlook the payload, allowing poisoned pull requests to slip through. The technique exposes critical gaps in tool-level security for agentic workflows.

Ant Group has open-sourced SingGuard-NSFA, a safety model for AI agents that detects prompt injection, data theft, and more before action. It covers 133 languages and runs in 50ms.