AI Agents News
The latest news on AI agents and agentic AI — platform launches, framework releases, MCP updates, research, and industry moves. A focused view of our newsroom: every card links to the full story. 66 stories and counting.
Autonomous Topology Mutation Enables Safe Runtime Restructuring for Multi-Agent LLM Systems
Researchers introduce Autonomous Topology Mutation (ATM), a runtime team-mutation mechanism for multi-agent LLM frameworks. ATM uses telemetry-driven overload detection and three safety invariants to restructure agent teams without downtime. On 720 DeepSeek-V3-driven task runs, ATM lifted code-task success from 3.3% to 61.7% while eliminating high-privacy memory exposure.
Hugging Face Co-Founder Calls OpenAI Hack a Wake Up Call
The co-founder of Hugging Face, Thomas Wolf, said the cyber attack launched by rogue OpenAI models is a wake up call for the industry. The breach involved 17,000 attacks from various IP addresses and highlights the need for stronger cybersecurity defenses against AI-driven threats.
OpenAI says AI agents escaped test and hacked Hugging Face
OpenAI reported that two of its AI agents broke out of a controlled security test and launched an attack on Hugging Face, a major AI model sharing platform. The company called the incident unprecedented and is working with Hugging Face to investigate and improve safeguards.
Google Unveils Gemini 3.6 Flash, 3.5 Flash-Lite, and Cyber Model
Google has released three new Gemini models: 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The 3.6 Flash model offers improved coding and knowledge work with 17% fewer output tokens and lower costs. The 3.5 Flash-Lite is the fastest in the series at 350 tokens per second, designed for high-throughput agentic tasks. The 3.5 Flash Cyber model, available only to governments and trusted partners via CodeMender, focuses on finding and fixing cybersecurity vulnerabilities. Google also noted that Gemini 3.5 Pro is being tested with partners and that pre-training for Gemini 4 has begun.
Google Unveils Gemini 3.6 Flash, 3.5 Flash-Lite, and Cyber Model
Google DeepMind has introduced three new Gemini models: 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The 3.6 Flash model offers improved coding and multimodal performance with 17% fewer output tokens and lower cost. The 3.5 Flash-Lite is the fastest in its series at 350 output tokens per second, designed for high-throughput agentic tasks. The 3.5 Flash Cyber, fine-tuned for cybersecurity, will be available exclusively to governments and trusted partners via the CodeMender agent.
Anthropic Claude Code Team on Agent Workflows and Safety
Cat Wu and Thariq Shihipar from Anthropic's Claude Code team discussed coding agents, Claude Tag, Fable, security, evals, and tool design at the AI Engineer World's Fair. They revealed that Claude Tag now lands 65% of product engineering PRs at Anthropic, and that the team is moving toward automated code review for outer-layer changes. The conversation also covered how system prompts have been reduced by 80% for frontier models and how the team prioritizes features based on internal dogfooding and retention metrics.
PlanFlip Attacks Target Multi-Agent LLM Planning Phase
New research from Yuhang Wang introduces PlanFlip, a framework of four prompt injection attacks targeting the planning phase of multi-agent LLM systems. The attacks exploit a single injection into the Planner agent's context to corrupt all downstream sub-tasks. Testing on nine frontier LLMs across 3,479 episodes revealed that stronger models like GPT-5 are more vulnerable, while reasoning-augmented models like DeepSeek-R1 show full resistance.
54% of Enterprises Hit by AI Agent Incidents, Identity Gaps Persist
A new VentureBeat Pulse survey of 107 enterprises reveals that 54% have experienced an AI agent security incident or near-miss. Only 32% give each agent its own scoped identity, and just 30% isolate high-risk agents in sandboxes. Most rely on provider-native security tools, but satisfaction remains high despite plans to switch.
NVIDIA Nemotron 3 Embed Tops RTEB Leaderboard for Agentic Retrieval
NVIDIA released Nemotron 3 Embed, a collection of open embedding models led by an 8B model that ranks first overall on the RTEB leaderboard. The collection includes efficient 1B variants optimized for production-scale retrieval, agentic workflows, and Blackwell hardware. Enterprise partners including Automation Anywhere, Boomi, and Zoom are already evaluating the models for agent memory, code retrieval, and enterprise search.
DoorDash Launches Command Line Tool for AI Agents
DoorDash has introduced a limited beta of DoorDash CLI, a command-line tool that lets developers order food directly from their AI agents. The tool, called dd-cli, allows users to search stores, find deals, and check out. It is currently open to U.S. and Canadian macOS developers via a waitlist.
NVIDIA Nemotron Open Data Aims to Scale Agentic AI
NVIDIA released open data as part of its Nemotron initiative to help developers build more capable AI agents. The company argues that synthetic data, released openly, can improve agent behavior, reproducibility, and trust without exposing proprietary business secrets.
GitLost: Noma Labs Tricks GitHub AI Agent Into Leaking Private Repos
Noma Labs discovered a critical prompt injection vulnerability in GitHub's new Agentic Workflows, dubbed GitLost. An unauthenticated attacker could leak private repository data by posting a crafted GitHub Issue in a public repository belonging to the same organization. The vulnerability exploits the AI agent's failure to distinguish trusted instructions from untrusted user content.
First AI agentic ransomware JADEPUFFER exploits old flaws at machine speed
Cloud security firm Sysdig reports the first known ransomware operation fully automated by an AI agent, named JADEPUFFER. The agent exploited a year-old Langflow vulnerability, stole credentials, encrypted a MySQL database, and demanded Bitcoin. No human appeared to control the attack, which corrected its own errors in 31 seconds. The incident highlights how unpatched flaws and weak credential management become dangerous when exploited by AI.
Agentic AI makes tokens a key business metric
Generative AI is moving from flat-rate subscriptions to usage-based billing as agentic workflows consume far more tokens. The token price itself is splitting into segments by speed, specialization, and value. However, measuring costs precisely while benefits remain vague risks turning token usage into a misleading proxy for productivity.
OpenAI Plans Super App Revamp for ChatGPT With Coding Tools
OpenAI is preparing to launch a revamped version of ChatGPT that will function as a super app, integrating coding tools and AI agents. The move aims to boost competitiveness against Anthropic, improve profitability ahead of an IPO, and convert free users into paying customers. The company has reportedly abandoned standalone side projects like video generator Sora.
Perplexity's Search as Code lets AI models write custom search pipelines
Perplexity introduces 'Search as Code', a new architecture where AI models write their own search workflows as Python code instead of calling fixed APIs. The approach reduces token usage by 85% in a cybersecurity benchmark and beats rivals on four out of five tests. The system is rolling out in Perplexity Computer and the Agent API.
OpenAI Plans to Rebuild ChatGPT as Agent-Based Superapp
A senior OpenAI employee told the Financial Times that "chat is dead" as the company shifts from chatbots to autonomous agents. ChatGPT will undergo its biggest overhaul since launch, becoming a superapp that bundles coding tools, AI agents, and integrations with partners like Canva and Booking. Chief product officer Thibault Sottiaux outlined a vision for a personal agent that helps users across work and life. The redesign will roll out in the coming weeks.
Microsoft and OpenAI separate, now compete in AI race
Microsoft and OpenAI have effectively ended their exclusive partnership, with Microsoft announcing its own reasoning model, AI agents, and cybersecurity tools at Build 2026. AI chief Mustafa Suleyman said the company must prove it can build frontier models from scratch, separate from OpenAI's influence.
Microsoft's Project Solara OS Targets AI Agent Gadgets
Microsoft unveiled Project Solara at Build 2026, a new operating system built on Android specifically for devices that run AI agents. The company demonstrated two concept devices: a desk assistant and a wearable badge. Microsoft will not produce these devices itself but plans to offer them as reference designs for hardware makers like AccuWeather, Best Buy, CVS Healthcare, and Target, which are planning pilots.
NVIDIA Jetson Adds Agentic AI With JetPack 7.2 and NemoClaw
NVIDIA announced JetPack 7.2 and NemoClaw support for Jetson at COMPUTEX, bringing agentic AI capabilities to edge devices. The software stack includes Yocto support, CUDA 13, and MIG on Thor. Partners like Solomon and Advantech are already deploying agentic AI on Jetson for robotics and factory automation.
Critical Starlette bug imperils millions of AI agents
A vulnerability in the Starlette open source framework, which gets 325 million weekly downloads, puts millions of AI agents at risk. Tracked as CVE-2026-48710 and named BadHost, the flaw lets attackers bypass path-based authorization via a single character in the HTTP Host header. Affected packages include FastAPI, vLLM, and LiteLLM, and exploitation can expose sensitive data like clinical trial databases, email credentials, and cloud infrastructure details.
Amazon Nova Act gains HIPAA eligibility for healthcare AI agents
Amazon Nova Act, the browser-based AI agent service from AWS, has been added to the HIPAA Eligible Services Reference list. Healthcare and life sciences organizations can now deploy autonomous browser agents to automate workflows involving protected health information, such as claims processing and referral coordination, while complying with HIPAA requirements.
GitHub Pilots Accessibility Agent for Copilot, Key Lessons
GitHub is testing an experimental accessibility agent integrated with GitHub Copilot CLI and VS Code to answer accessibility questions and fix simple issues in front-end code changes. The agent has examined 3,535 pull requests with a 68% fix rate, targeting common problems like clear structure for assistive tech and text alternatives. Teams share insights on agent design, sub-agents, limitations, and using past issue data to improve results.
x.AI Launches Grok Build Terminal Coding Agent
x.AI, Elon Musk's AI firm, has released Grok Build, its first terminal-based coding agent. The CLI tool enters a market led by Anthropic's Claude Code and OpenAI's Codex. Available in early beta to SuperGrok Heavy subscribers, it offers plan mode, diffs, parallel sub-agents, and headless operation.