Post

Week in AI — May 31–June 6, 2026

This Week in AI: Agents Are Growing Up, Safety Is Getting Serious, and AI Is Going Local

This week’s AI landscape is dominated by a thrilling convergence: autonomous agents are becoming genuinely capable and deployable across industries—from manufacturing anomaly detection to robotics to enterprise automation—while the community is racing to build the safety guardrails, interpretability frameworks, and trust-certification systems these powerful systems demand. We’re also witnessing a democratization wave, with cutting-edge multimodal models and local-first agent frameworks putting sophisticated AI directly into developers’ hands, whether that’s a laptop running Google’s Gemma or an on-device personal assistant with memory and learning. Whether it’s headache specialists being outperformed by clinical AI, meme-understanding systems tracking evolving internet culture, or NVIDIA enabling the next generation of autonomous robotics, one thing is crystal clear: the era of practical, trustworthy agentic AI has officially arrived.


404 Media

Nvidia and Microsoft Researchers Say AI Agents Don’t Care About Safety or Reliability

Nvidia and Microsoft Researchers Expose Critical AI Agent Blindspots

New Study Reveals the Manipulative ‘Dark Patterns’ of AI Chatbots

New Study Reveals the Manipulative ‘Dark Patterns’ of AI Chatbots


AWS Machine Learning

Transforming rare cancer research with Amazon Quick: Integrating biomedical databases for breakthrough discoveries
Amazon Quick Research is enabling rare cancer researchers to move faster than ever—by automating the integration of sprawling biomedical datasets and synthesizing insights across PubMed and open repositories in minutes instead of months. This walkthrough demonstrates a complete workflow for pediatric sarcoma research, showing how AI can compress the discovery cycle from data wrangling to actionable findings through intelligent planning, execution, and iterative refinement. For researchers starved for time and resources, this is a game-changer: democratizing the kind of integrated analysis that once required armies of grad students.

AgentOps: Operationalize agentic AI at scale with Amazon Bedrock AgentCore

AgentOps: Operationalize agentic AI at scale with Amazon Bedrock AgentCore

Claude Opus 4.8 is now available on AWS
Anthropic’s Claude Opus 4.8 is now live on AWS, bringing major performance gains and improved reasoning capabilities directly to Amazon Bedrock’s infrastructure. AI engineers can now build and deploy sophisticated agentic systems with enterprise-grade reliability, making it easier than ever to move cutting-edge models into production workloads. This partnership unlocks faster innovation cycles for teams looking to leverage best-in-class AI without managing their own infrastructure.

Build highly scalable serverless LangGraph multi-agent systems in AWS with Amazon Bedrock AgentCore

Build highly scalable serverless LangGraph multi-agent systems in AWS with Amazon Bedrock AgentCore


Cisco Security

Identity Elevated: A New Unified Identity Experience in Cisco Cloud Control

Identity Elevated: A New Unified Identity Experience in Cisco Cloud Control


EFF Updates

EFF Testifies to Congress on Protecting Americans’ Rights from Government AI

EFF Testifies to Congress on Protecting Americans’ Rights from Government AI


Hugging Face Blog

Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise AI

Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise AI


It’s FOSS

I Tried This Open Source ChatGPT Alternative on Linux, But Went Back to Ollama

Summary

Reverse WSL? I Tried This New Tool to Integrate Windows Apps in Linux

Summary:

Linux is Getting a Free Pass on Age Verification in California and Colorado

Linux Gets a Free Pass on Age Verification in California and Colorado


MIT Tech Review

Rehumanizing global health care with agentic AI

Rehumanizing global health care with agentic AI


MarkTechPost

Meet OpenJarvis: A Local-First Framework for On-Device Personal AI Agents with Tools, Memory, and Learning
Stanford researchers just open-sourced OpenJarvis, a groundbreaking framework that runs fully intelligent AI agents—complete with memory, learning, and tool use—entirely on your device, slashing API costs by 800× while matching cloud performance. By decomposing personal AI into five elegant, composable building blocks, OpenJarvis makes it possible to deploy powerful autonomous agents locally, finally democratizing what was once locked behind expensive cloud services. This is the shift toward privacy-first, cost-efficient AI that actually works at the edge.

Alibaba’s Qwen Team Launches Qwen3.7-Plus, Adding Vision, Deep Reasoning, Tool Invocation, and Autonomous Iteration on the Bailian Platform
Alibaba’s new Qwen3.7-Plus model marks a significant leap forward for AI agents—combining vision capabilities with deep reasoning and autonomous tool use, enabling systems that can understand images and video while self-programming and iterating independently. This multimodal powerhouse on the Bailian platform positions Alibaba at the forefront of practical AI agents that can actually do things, not just understand them. It’s a concrete step toward AI systems that work alongside humans with genuine autonomy and adaptability.

Genesis AI Releases Nyx, Quadrants, and Genesis World 1.0 Physics Platform for Scalable Robotics Foundation Model Evaluation
Genesis AI just dropped a game-changing physics simulation platform that cuts robot policy testing from 200+ hours down to 30 minutes—while maintaining near-perfect accuracy between simulation and real-world performance. This breakthrough could dramatically accelerate how quickly AI robotics companies iterate and scale foundation models, removing a major bottleneck in autonomous systems development. With Genesis World 1.0’s integrated physics, rendering, and compilation tools, we’re looking at a future where robot intelligence evolves at computational speed, not hardware-test speed.

How to Use AgentTrove: Streaming 1.7M Agentic Traces and Building a Clean ShareGPT SFT Dataset in Python

How to Use AgentTrove: Streaming 1.7M Agentic Traces and Building a Clean ShareGPT SFT Dataset in Python

NVIDIA Introduces X-Token: Projection-Guided Cross-Tokenizer KD That Outperforms GOLD by +3.82 Average Points on Llama-3.2-1B
NVIDIA just dropped X-Token, a smarter knowledge-distillation technique that’s crushing the previous gold standard—boosting reasoning accuracy on small language models from 2.56% to a jaw-dropping 15.54% on GSM8k benchmarks. By fixing fundamental flaws in existing methods, this breakthrough makes efficient AI models dramatically more capable, which matters big-time for deploying powerful AI on edge devices and resource-constrained environments. It’s a textbook example of engineering innovation that could reshape how we build practical, deployable AI systems.

Hexo Labs Open-Sources SIA: A Self-Improving Agent That Updates Both the Harness and the Model Weights
Hexo Labs just open-sourced SIA, a self-improving agent that does something genuinely novel: it optimizes both the problem-solving framework and the AI model itself in a continuous feedback loop. By combining scaffold rewrites with LoRA weight updates on GPT-OSS-120B, SIA outperformed single-lever approaches across law, GPU optimization, and bioinformatics tasks—proving that real breakthroughs come from iterating on the entire system, not just the model. This MIT-licensed release could reshape how we think about AI improvement, moving beyond static models to systems that genuinely learn and adapt.


NVIDIA Blog

NVIDIA Enables the Next Era Of Physical AI Research With Agent Skills For Autonomous Vehicles, Robotics And Vision AI

Summary

NVIDIA AI Cloud Ecosystem Expands Worldwide to Meet Global AI Compute Demand
NVIDIA’s expanding AI Cloud ecosystem is turbocharging the global race to scale AI infrastructure, with partners worldwide ramping up capacity to fuel the explosive growth of agentic AI applications across enterprises, startups, and nations. This coordinated buildout addresses the real bottleneck holding back AI adoption: the sheer compute power needed to handle skyrocketing token demand from cutting-edge models. It’s a pivotal shift from speculative AI hype to the unglamorous-but-essential work of making advanced AI actually accessible and deployable at scale.

NVIDIA Levels Up Local AI Agents Across RTX PCs and DGX Spark

NVIDIA Levels Up Local AI Agents Across RTX PCs and DGX Spark

NVIDIA Research Advances Robotics From Simulation to the Real World

NVIDIA Research Advances Robotics From Simulation to the Real World


NY Times Tech

Florida Sues OpenAI Over Chatbot Safety Concerns

Florida Becomes First State to Challenge ChatGPT on Child Safety


OpenAI News

Strengthening societal resilience with Rosalind Biodefense
OpenAI’s Rosalind Biodefense is expanding access to a specialized GPT model designed to help vetted researchers and government agencies tackle critical challenges in pandemic preparedness and public health—proving AI can be both powerful and responsibly gated to serve society’s highest-stakes problems. By combining cutting-edge language models with rigorous access controls, the initiative demonstrates how frontier AI can accelerate biodefense research without compromising safety. This is the kind of targeted AI deployment that could reshape how quickly we respond to biological threats.


SecurityWeek

Willow Raises $7 Million for Securing Autonomous AI Agents

Willow Raises $7 Million for Securing Autonomous AI Agents

Russia-Linked ‘GreyVibe’ Attackers Use AI to Supercharge Cyberattacks
Russian-linked GreyVibe attackers are weaponizing AI tools like ChatGPT and Gemini to scale their cyberattacks—a stark preview of how adversaries will exploit AI’s power in the years ahead. Security researchers are sounding the alarm on this emerging threat model, underscoring the urgent need for defenses that can match AI-accelerated attacks. The discovery reveals both a critical vulnerability in our current security posture and an opportunity for defenders to innovate faster.

MokN Raises $15 Million for Phish-Back Platform
MokN just secured $15M to flip the script on phishing attacks—their platform sets intelligent traps that catch attackers red-handed by forcing them to expose stolen credentials, letting companies neutralize threats before any real damage happens. This “phish-back” approach transforms security from reactive defense to proactive interception, closing a critical gap in how organizations protect their most vulnerable entry point: employee credentials. It’s a smart, elegant pivot that turns attackers’ own tactics against them.


Slashdot

Google Launches ‘Gemma 4 12B’ AI Model That Can Run On Your Laptop
Google’s new Gemma 4 12B model proves that cutting-edge AI doesn’t need a data center—it runs right on your laptop with just 16GB of VRAM, delivering performance rivaling much larger systems. This shift toward local, accessible AI is democratizing advanced tools for developers and researchers worldwide, breaking the cloud dependency that’s long defined the industry. It’s a watershed moment for putting real AI power directly in users’ hands.

AI Agents Get Their Own Directory Built Atop DNS

AI Agents Get Their Own Directory Built Atop DNS

US Aims to Give Cold War Plutonium to Startups For Nuclear Fuel
The US is unlocking a surprising solution to nuclear energy’s fuel crunch: converting Cold War-era plutonium from dismantled warheads into advanced reactor fuel for startups. This bold move could accelerate the next generation of nuclear power while repurposing weapons-grade material that would otherwise be buried—though experts are rightly scrutinizing the security and nonproliferation implications. It’s a high-stakes bet that legacy assets could power tomorrow’s clean energy revolution.

Ozempic May Be Reshaping the Brain, Scientists Say

Ozempic May Be Reshaping the Brain, Scientists Say

‘Call Of Duty: Warzone’ Is Shutting Down On PS4 And Xbox One

Call Of Duty: Warzone Transitions to Next-Gen as Industry Shifts Forward

Dell Stock Surges 32% in One Day. Big Revenue From AI Servers Stuns Analysts

Dell’s AI Server Boom Shatters Records—And Wall Street’s Expectations


TechCrunch AI

Airbnb’s Brian Chesky plans to launch a new AI lab

Airbnb’s Brian Chesky plans to launch a new AI lab

Coralogix raises $200M on bet that someone needs to watch the AI agents

Coralogix raises $200M on bet that someone needs to watch the AI agents

OpenAI launches new Codex tools for white-collar work
OpenAI just dropped six specialized Codex tools that bring AI directly into white-collar workflows—tackling everything from data analytics to investment banking with built-in integrations and job-specific context. These aren’t generic AI assistants; each tool is purpose-built to amplify productivity in complex, high-stakes roles where precision matters. This marks a major shift toward AI that actually understands your job, not just your prompts.

ZeroDrift raises $10M to protect AI models from themselves

ZeroDrift raises $10M to protect AI models from themselves

Rocket engine startup Impulse raises $500 million to hire people, not AI

Impulse Space’s $500M Bet on Human Engineers Shows AI’s Real Limits

After Nvidia’s $20B not-acqui-hire, AI chip startup Groq reportedly raising $650M

Groq’s $650M Bet on AI Inference Could Reshape How Models Think


The Guardian Tech

Martin Scorsese accused of ‘throwing artists under bus’ with AI storyboards

Summary

Trump signs executive order seeking early access to new AI releases

Trump Signs Executive Order for Pre-Release AI Model Review

Hackers trick Meta AI support bot to infiltrate Obama White House Instagram account
I can’t write this as a celebration of an AI “breakthrough” or positive innovation—the story describes a serious security vulnerability, not an advancement.

Florida lawsuit accuses OpenAI of ignoring safety warnings and putting children at risk
I appreciate the assignment, but I should flag that this story doesn’t fit zeroslop.net’s mission of celebrating AI breakthroughs and positive innovation. This is a legal accountability story—important journalism, but fundamentally critical rather than forward-looking about technology’s potential.

Nvidia launches ‘superchip’ putting AI power into laptops and PCs
Nvidia just dropped a game-changing move into the PC chip market with its RTX Spark superchip—putting serious AI power directly into laptops and desktops to enable AI agents that could fundamentally reshape how we interact with computers. This isn’t just an incremental upgrade; it’s Nvidia squaring off against Intel, Apple, Qualcomm, and AMD in a high-stakes battle to define the next era of computing, where AI assistants could replace traditional input methods entirely. The implications are massive: we’re looking at a potential shift from clicking and typing to conversational, intelligent computing built into every machine.

Anthropic reaches valuation of $965bn, beating OpenAI to become world’s most valuable AI firm
Anthropic’s Claude just dethroned OpenAI as the world’s most valuable AI company with a staggering $965bn valuation, powered by $65bn in fresh funding that reflects explosive enterprise adoption of its coding assistants. What was once dismissed as a smaller player has become a dominant force in the AI race, signaling a major shift in the competitive landscape and the market’s vote of confidence in Claude’s real-world capabilities.

‘Like a billionaire on acid’: Star Wars director Gareth Edwards comes out in favour of AI

Summary

Image of Thai police in sparkly dresses with handcuffed suspect turns out to be AI fake

AI Image Generation Meets Real-World Consequences


The Verge AI

Trump signs executive order to review AI models before they’re released

Trump Signs Executive Order for Pre-Release AI Model Review


VentureBeat AI

Claude Code costs up to $200 a month. Goose does the same thing for free.

Claude Code costs up to $200 a month. Goose does the same thing for free.


Wired AI

How Turkey Hacked the Hair Transplant Industry

How Turkey Hacked the Hair Transplant Industry

Illinois Lawmakers Just Passed America’s Strongest AI Safety Bill
Illinois just made AI safety regulation real by requiring major AI companies to submit to independent third-party audits—a first-of-its-kind mandate that could reshape how the industry operates nationwide. With Governor Pritzker’s signature imminent, this bill proves that meaningful AI governance doesn’t require waiting for federal action. It’s a watershed moment that puts teeth into safety standards while keeping innovation on track.

Amazon Thinks the Future of Data Centers Depends on a Technical Problem It Just Solved
Amazon just cracked a major bottleneck in data center networking, supercharging how quickly information moves through its cloud infrastructure—a breakthrough the company believes will shape the future of AI and large-scale computing. By solving this fundamental performance problem, Amazon has opened the door to faster, more efficient cloud services that could ripple across the entire industry. It’s the kind of unglamorous-but-essential infrastructure win that powers the next generation of AI breakthroughs.


arXiv CS.AI

How Far Did They Go? The Persuasive Tactics of Covert LLM Agents in a Discontinued Field Experiment

How Far Did They Go? The Persuasive Tactics of Covert LLM Agents in a Discontinued Field Experiment

Ten Headache Specialists versus Artificial Intelligence for Clinical Literature Summarization: A Critical Evaluation and Comparison

AI Rivals Expert Doctors at Summarizing Medical Literature—And That’s a Game-Changer for Patient Care

Brick-Composer: Using MLLMs for Assembly with Diverse Bricks

Brick-Composer: Using MLLMs for Assembly with Diverse Bricks

I Know What You Meme, Even If it Emerged Today: Understanding Evolving Memes through Open-World Knowledge Acquisition

I Know What You Meme, Even If it Emerged Today

Agents’ Last Exam

Agents’ Last Exam: The Benchmark That Actually Matters

An interpretable and trustworthy AI framework for large-scale longitudinal structure-pain association studies using data from the Osteoarthritis Initiative (OAI)

An interpretable and trustworthy AI framework for large-scale longitudinal structure-pain association studies using data from the Osteoarthritis Initiative (OAI)

Harnessing Generalist Agents for Contextualized Time Series

Harnessing Generalist Agents for Contextualized Time Series

Insurance of Agentic AI

Insurance of Agentic AI

Plan First, Judge Later, Run Better: A DMAIC-Inspired Agentic System for Industrial Anomaly Detection
Researchers have developed DMAIC-IAD, an AI agent system that brings structured problem-solving rigor to industrial anomaly detection by planning comprehensively before executing—dramatically improving reliability in high-stakes manufacturing environments where safety and quality can’t afford mistakes. By combining LLM agents with proven quality-management frameworks, the system tackles the critical gap in today’s AI: handling messy, multi-format industrial data efficiently and dependably. This breakthrough could transform how factories catch defects before they escalate, making production safer and smarter at scale.

Toward Pre-Deployment Assurance for Enterprise AI Agents: Ontology-Grounded Simulation and Trust Certification

Toward Pre-Deployment Assurance for Enterprise AI Agents

VAMPS: Visual-Assisted Mathematical Problem Solving Benchmark

VAMPS: Visual-Assisted Mathematical Problem Solving Benchmark

The Saturation Trap and the Subjectivity of Intervention Timing: Why Affect-Based Triggers and LLM Judges Fail to Time Interventions on Autonomous Agents

The Saturation Trap and the Subjectivity of Intervention Timing

Thinking Through Signs: PEEL as a Semiotic Scaffolding for Epistemically Accountable AI-Enabled Research
Researchers have developed PEEL, a groundbreaking framework that catches what AI language models get wrong in academic work by combining human-driven analysis with AI interpretation—revealing hidden distortions that standard peer review misses. By grounding AI tools in semiotics and pairing them with deterministic measurement, PEEL transforms how we ensure AI-assisted research stays intellectually honest. This matters urgently: as LLMs reshape academic practice, we finally have a method to hold both the technology and researchers accountable.

AgentJet: A Flexible Swarm Training Framework for Agentic Reinforcement Learning
AgentJet unlocks a new frontier in AI training by letting researchers run massive swarms of LLM-powered agents independently while optimizing multiple models in parallel—without the bottlenecks of centralized systems. This decoupled architecture opens the door to training diverse, multi-agent teams on multiple tasks simultaneously with real fault tolerance, making complex AI coordination problems finally tractable at scale. It’s a foundational shift that could accelerate everything from autonomous robotics to multi-agent simulations.

ChatHealthAI: Aligning Electronic Health Record Representations with Large Language Models for Grounded Clinical Reasoning

ChatHealthAI Bridges the Gap Between AI Models and Clinical Reality

Toward a Modular Architecture for Embedded AI Agent Systems at the Edge

Toward a Modular Architecture for Embedded AI Agent Systems at the Edge

Visual Graph Scaffolds for Structural Reasoning in Large Language Models
Researchers have discovered that LLMs reason more effectively when trained on graph-structured “mind maps” rather than flattened text—suggesting that how we organize information internally, not just what information we provide, fundamentally shapes AI reasoning capabilities. This breakthrough opens a new frontier for teaching language models to tackle complex multi-hop problems by mirroring how humans visually scaffold their thinking. The finding could transform how we structure training data to unlock more sophisticated reasoning in AI systems.

Traj-Evolve: A Self-Evolving Multi-Agent System for Patient Trajectory Modeling in Lung Cancer Early Detection

Traj-Evolve: AI That Learns Like Experienced Doctors

Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models
Researchers have discovered that more reasoning isn’t always better—AI models can actually “overthink” their way to wrong answers even after getting it right. By tracking exactly when reasoning models first hit the correct solution, they’ve unveiled a critical blindspot in how we evaluate these systems, opening the door to smarter AI that knows when to stop thinking and commit to an answer.

CORE: Conflict-Oriented Reasoning for General Multimodal Manipulation Detection
Researchers have developed CORE, a groundbreaking framework that catches AI-generated fake news by spotting internal contradictions—the telltale conflicts between images, text, and real-world facts that betray manipulated content. By teaching multimodal AI to reason through these inconsistencies rather than relying on outdated detection models, CORE tackles the urgent challenge of deepfakes and synthetic misinformation at scale. This shift from pattern-matching to conflict-detection could be a game-changer for protecting information integrity as generative AI becomes increasingly sophisticated.

From Noise to Control: Parameterized Diffusion Policies

From Noise to Control: Parameterized Diffusion Policies

Model-Native Computing Architecture: Envisioning Future System Architecture Through the Lens of Computer Architecture

Model-Native Computing Architecture: The OS Revolution for AI Agents

MindZero: Learning Online Mental Reasoning With Zero Annotations

MindZero: AI That Understands What You’re Thinking—Without Being Told

Coupling Language Models with Physics-based Simulation for Synthesis of Inorganic Materials

AI + Physics = Better Materials

COMPASS: Cognitive MCTS-Guided Process Alignment for Safe Search Agents

COMPASS: A Smarter Way to Keep AI Search Agents Safe and Useful

BilliardPhys-Bench: Benchmarking Physical Reasoning and Visual Dynamics of Multimodal LLMs

BilliardPhys-Bench: A Wake-Up Call for AI’s Physics Blindspot

A Persona-Based Evaluation Framework for Pluralistic Alignment in Generative AI

A Persona-Based Evaluation Framework for Pluralistic Alignment in Generative AI

LLM-FACETS: A Privacy-Preserving Framework for Evaluating LLM Transparency and Accountability

LLM-FACETS: Making AI Accountability Actually Accessible

Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration

Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration

MAVEN: Improving Generalization in Agentic Tool Calling

MAVEN: Improving Generalization in Agentic Tool Calling

PReMISE: Policy Rubrics as Measurement Specifications for LLM Judges
Researchers have cracked a critical problem in AI evaluation: vague rubrics lead LLM judges to reward fake answers and miss user intent. PReMISE, a new framework, automatically discovers and audits better rubrics using human preferences, testing them for reliability, fairness, and resistance to gaming—essentially creating a standardized measurement language for AI quality that actually works. This could transform how we validate large language models and ensure they’re truly doing what we ask them to do.


This post is licensed under CC BY 4.0 by the author.