Google overhauled Search at I/O 2026, replacing blue links with AI agents. The backlash has been swift. DuckDuckGo app installs spiked 30% as users seek a way out.

OpenRouter has raised a $113 million Series B led by CapitalG. Its 5x growth in usage over six months indicates the multi-AI-model future is here.

Human Archive, a startup founded by UC Berkeley and Stanford researchers, is paying gig workers in India to wear camera-equipped caps and sensor devices to collect the real-world physical training data that AI and robotics labs are racing to acquire.

For years, UMG has pushed platforms, streaming services, and AI companies to implement stricter content moderation policies.

Save up to $410 on your TechCrunch Disrupt 2026 pass before prices increase on May 29 at 11:59 p.m. PT. Register here to join the tech epicenter in San Francisco.

Autonomous AI systems are beginning to move beyond software environments and into warehouses, delivery networks, and public spaces. The development is drawing attention to whether current AI rules cover systems that operate in physical environments.

Most existing AI governance frameworks have focused on online harms and model outputs, including bias, misinformation, and harmful content. Embodied AI systems carry risks in physical environments, where failures can affect infrastructure, property, or human safety.

Singapore’s Infocomm Media Development Authority published version 1.5 of its Model AI Governance Framework for Agentic AI on May 20. The framework sets out guidance for organisations deploying AI agents that can plan, make decisions, and take actions across multiple steps to complete user-defined goals.

The framework says agents can interact with tools, external systems, and other agents, including systems that update databases, write files, control devices, or perform transactions. It lists access controls, monitoring, and human approval among governance measures for deployment.

AI moves into physical systems

At an AI summit in Singapore last week, discussions around robotics and embodied AI focused on operational safety issues more commonly associated with aviation, industrial systems, and critical infrastructure oversight than conventional software regulation.

Speakers also discussed whether autonomous systems can operate safely and reliably in unpredictable real-world environments over extended periods.

Dr. Ya-Qin Zhang, founding dean of the Institute for AI Industry Research at Tsinghua University, said embodied AI systems amplify risks already associated with autonomous software. He said failures can directly affect transport systems, drones, logistics networks, and critical infrastructure.

“Any risk in the digital domain will be amplified in the physical domain, and the physical domain will have a physical consequence,” Zhang told MLex on the sidelines of the summit.

He added that vehicles, drones, smart grids, and other infrastructure could become exposed as AI systems are embedded more deeply into physical operations.

Speakers discussed reliability, operational monitoring, and post-deployment assurance as governance concerns. Summit discussions pointed to deployment-based governance models built around simulation, telemetry, and iterative testing, rather than one-time certification alone.

IMDA’s framework also recommends gradual rollouts, continuous monitoring, and further testing after deployment. It says agents interact dynamically with their environment and not all risks can be anticipated before release.

Monitoring becomes a deployment issue

Grab, which is piloting autonomous vehicles and delivery robots in Singapore’s Punggol district, said deployment governance depends heavily on simulation, testing, and continuous monitoring.

“We do a lot of simulation, we do a lot of testing in closed courses and open courses in order to make sure our robots are reliable,” Suthen Thomas Paradatheth, Grab’s chief technology officer, said during one of the summit panels.

“Before we scale to hundreds of robots, we make sure we crack it first in simulation and with a few robots,” he added.

Grab also pointed to monitoring systems designed to track robot performance and detect unexpected failures after deployment.

“There’s a long tail of issues that could emerge,” Paradatheth said.

The IMDA framework says organisations should assess agentic AI use cases based on data access, external system access, autonomy, and task complexity. It also points to the scope and reversibility of agent actions, third-party involvement, and overall system complexity.

It also recommends limiting agent access to tools and systems, applying least-privilege permissions, and defining standard operating procedures for agent workflows. Organisations should also set mechanisms to take agents offline when they malfunction.

Accountability spreads across more actors

MLex reported that embodied AI systems can involve several parties across development, manufacturing, and deployment. These include AI developers, robotics manufacturers, semiconductor suppliers, and infrastructure operators.

MLex also noted that responsibility can be harder to assign when systems continue adapting after deployment through software updates, telemetry, and operational data.

IMDA says organisations and humans remain accountable for agent actions, even when agents operate autonomously. The framework calls for clear responsibility across the agentic AI value chain, from model and platform providers to deployers, tooling providers, and end users.

Applied Materials said large-scale robotics deployment is also tied to semiconductor economics and systems integration. Om Nalamasu, the company’s chief technology officer, said robotics systems will depend on better sensors, energy efficiency, advanced packaging, and computing architectures.

Nalamasu said robotics systems would require purpose-built designs adapted to specific industrial ecosystems rather than a single solution for all environments.

Zhao Yuli, chief strategy officer of Chinese robotics startup Galbot, said Beijing is prioritising deployment scale and industrial commercialisation through government-backed testbeds, industrial partnerships, and long-term funding initiatives.

Galbot has deployed humanoid robotics systems in retail, warehouse, and pharmaceutical operations in China. These include autonomous stores that operate around the clock. Zhao said semi-structured industrial environments are likely to become an early commercialisation path because they offer more controllable operating conditions.

Japan is placing more focus on standards-setting, robotics datasets, and safety governance. Professor Yutaka Matsuo of the University of Tokyo’s Graduate School of Engineering pointed to an “AI Association” project aimed at collecting 100,000 hours of robotics data to support robotic foundation models.

Matsuo also referred to Japan’s AI Safety Institute and the Hiroshima AI Process as part of broader efforts to develop governance standards for embodied AI systems with Singapore and other Asian countries.

Singapore sets out agent controls

Singapore’s framework sets out four governance areas for agentic AI. These cover upfront risk assessment, human accountability, technical controls, and end-user responsibility. The framework describes them as an iterative process rather than a one-time assessment.

The framework says human oversight has to be adapted for agentic systems because continuous review of all workflows becomes impractical at scale. It recommends human approval at significant checkpoints, including high-stakes actions, irreversible actions, and outlier behaviour.

IMDA also identifies automation bias and alert fatigue as risks when humans supervise capable agents. It recommends auditing oversight through indicators such as human override rates and response times, and using automated real-time monitoring to flag unexpected behaviour.

The framework says users should be told what actions an agent can take, what data it can access, and what responsibilities remain with the user. It also recommends employee training on human-agent interaction, oversight, and the professional skills needed to assess agent outputs.

Companies test AI in regulated workflows

JPMorgan is implementing AI tools across its global investment banking business, Paul Uren, the bank’s Asia Pacific head of investment banking, told Reuters. The bank said the tools help bankers access more information and synthesise it with internal systems. They are also being used to prepare content and support client engagement.

JPMorgan CEO Jamie Dimon told Bloomberg News that the bank would hire more AI specialists and fewer traditional bankers. Reuters reported that global banks are increasing AI investment, reshaping workforces, and changing job roles.

The bank is also among selected organisations permitted by Anthropic to use its Mythos cybersecurity model under a controlled initiative known as Project Glasswing. According to Anthropic, Mythos can detect old vulnerabilities in browsers, infrastructure, and software.

Reuters reported that Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley also have access to, or are testing, Mythos, citing sources and company executives.

IMDA’s framework includes a case study from OCBC Bank of Singapore on source-of-wealth analysis. The system parses income-related documents and drafts a source-of-wealth memo. It does not make credit, onboarding, or risk decisions autonomously.

In that case, the workflow is limited to task-level autonomy and operates only when triggered by predefined workflows. Human review is required at critical decision points, and final validation remains with designated reviewers.

Robots move into industrial use

In Japan, one-third of companies are already using or considering AI-powered robots, according to a Reuters survey conducted by Nikkei Research from May 1 to May 15. The survey contacted 492 companies, with 220 responding on the condition of anonymity.

About 4% of respondents said they already use AI robots, 5% plan to deploy them, and 25% are considering doing so. The remaining 66% said they had no such plans.

Transportation equipment manufacturers were the most active group in the survey, with 80% already using AI robots or considering deployment. By comparison, 94% of wholesale sector respondents said they had no plans to deploy AI robots.

Among companies using, planning to use, or considering AI robots, 71% selected manufacturing as a use case. Another 19% selected dangerous tasks, while 11% selected customer-facing services.

The Japanese government expects AI robots to help address the country’s chronic labour shortage and support its position in industrial robotics. Japan is home to robotics companies including Fanuc, Yaskawa Electric, and Kawasaki Heavy Industries, but faces competition from China and the United States in AI-enabled robotics.

Retail agents expand beyond search

Walmart has outlined plans to use agentic AI across shopping, employee, supplier, and developer workflows.

In July 2025, the retailer announced plans for four AI-powered “super agents.” They are designed for shoppers, store employees, suppliers and sellers, and software developers. Walmart said these agents would become the main entry point for AI interactions across those groups.

One of the tools, Sparky, is already available in Walmart’s app as a generative AI-powered shopping assistant. Hari Vasudev, Walmart’s US chief technology officer, said its expanded version would be able to reorder items and plan events. It would also use computer vision to suggest recipes based on the contents of a shopper’s fridge.

Walmart is also developing an Associate super agent for store workers and corporate staff. A separate Marty agent is being built for sellers, suppliers, and advertisers. The retailer is also working on a Developer super agent for testing, building, and launching future AI tools.

The company declined to say whether the agents would replace jobs. Dave Glick, senior vice president of enterprise business systems, said the tools would create new jobs, without giving further details.

(Photo by Growtika)

See also: OpenAI opens Singapore AI lab as IMDA updates AI framework

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.

The post Autonomous AI systems test governance in physical environments appeared first on AI News.

InfoQ expands its online certification portfolio with new AI Engineering and Organizational Architecture cohorts, giving senior practitioners a confidential peer group to pressure-test production AI, platform, team design, and architecture decisions.

By Artenisa Chatziou

This article explores Kafka's transition toward a cloud-native architecture, examining how tiered storage, FinOps telemetry, elastic consumer scaling, virtual clusters, and Share Groups reshape the operational and economic model of event streaming platforms. It also analyzes emerging diskless-storage proposals and their architectural trade-offs.

By Viquar Khan

Google's SynthID, designed to embed imperceptible signals into AI-generated content, is adding a new Content Detection API on Google Cloud's Gemini Enterprise Agent Platform, after gaining adoption by several industry players including Nvidia and OpenAI.

By Sergio De Simone

arXiv:2605.23908v1 Announce Type: new Abstract: We are in the midst of large-scale industrial and academic efforts to automate the processes of scientific, technological and creative production through AI-driven assistants. Historically, a fundamental property of these processes in their human form has been their open-endedness: their capacity for generating a seemingly endless supply of novel and meaningful new forms. Do artificial agents have any capacity for such fruitful unguided discovery? To answer this question, we turn to Picbreeder, the canonical exemplar of human-driven open-ended search, in which users collaboratively generated a diverse library of images through interactive evolution of small neural networks. We replicate Picbreeder, replacing human users with frontier Vision Language Models (VLMs). We observe clear qualitative differences between the output of our system and the historical human baseline, and attempt to characterize them using metrics of phylogenetic complexity and visual and semantic salience and novelty. In an effort to identify some of the causal factors contributing these differences, we study the addition of exploratory noise to the agents' selection process, of behavioral diversity between agents, and of narrative momentum in the form of memory of past actions. We make our code available at https://github.com/smearle/picbreeder-vlm.

arXiv:2605.23909v1 Announce Type: new Abstract: We investigate the calibration of large language models' (LLMs') confidence across diverse tasks. The results of our preregistered study show that the current crop of LLMs are, like people, too sure they are right: confidence exceeds accuracy, on average. Importantly, however, this tendency is moderated by a powerful hard-easy effect, wherein overconfidence is greatest on difficult tests; by contrast, easy tests actually show substantial underconfidence. We develop LifeEval, a test for evaluating model calibration across levels of difficulty.

arXiv:2605.23926v1 Announce Type: new Abstract: Reasoning-capable large language models solve hard problems by emitting long chains of thought, paying heavily in latency, GPU time, and energy. Casual inspection of their traces reveals extensive reformulation, verification, and circular self-reflection, yet how much of this deliberation is actually necessary has never been measured at scale or explained from first principles. This paper closes both gaps. We formalise reasoning redundancy directly in terms of the reasoning model itself: the redundancy of a correct trace is the largest fraction of its trailing segmented steps that can be truncated while $\pi$, forced to terminate thinking and emit a final answer, still produces the correct answer. A large-scale quantification across four frontier reasoning models and two mathematical benchmarks shows that step-level redundancy is consistently high -- between 61% and 93% across the 8 (model, benchmark) conditions we study, with the median critical prefix equal to a single segmented step in six of the eight conditions -- that the finding is robust to the choice of judge family, and that although $\rho$ decreases with problem difficulty on MATH-500, all four models remain substantially redundant ($\rho \in [46\%, 85\%]$) even on the hardest Level-5 problems. We then prove that this redundancy is a structural consequence of length-agnostic outcome rewards, not a model-specific artefact: under any such reward, no finite expected stopping time is optimal. The result holds regardless of RL algorithm, base model, data distribution, or whether the policy is obtained via RL or distillation; over-thinking is therefore not a bug to be patched in individual models but a structural property of how current reasoning models are trained. Code: https://github.com/zhiyuanZhai20/how-much-thinking-is-enough

arXiv:2605.23928v1 Announce Type: new Abstract: We present Context, the intelligence layer of the Magarshak Architecture, which replaces reactive query-response chatbots with proactive goal-directed agents that advance shared tasks without waiting for user prompts. The architecture rests on three mutually reinforcing mechanisms. Write-time context assembly precomputes enriched typed attributes via Groker agents, assembling interaction context as a deterministic pure function of graph state; context blocks are byte-identical across turns between semantic changes, enabling near-100% KV-cache reuse. Composable sandboxed wisdom programs form a governed library of LM-generated imperative programs declaratively wired to goal types via typed stream relations, composed via phase ordering, and executed at interaction time without further LM calls. Proactive goal stream state machines drive conversations toward terminal states by inspecting graph state and emitting structured interaction content (option arrays, governance affordances, clarification prompts) without awaiting user input. We prove six formal results: the Context Stability Theorem, bounding per-turn LM cost as a function of semantic change rate; a Program Composition Correctness Theorem; a Declarative Wiring Soundness Theorem; the Proactive Dominance Theorem, proving proactive agents weakly dominate reactive agents on expected turns-to-terminal-state; Coordination Overhead Elimination and Quality Preservation, establishing Pareto improvements in multi-participant goal chats; and a Cross-Platform Vote Consistency Theorem. Implemented in the open-source Qbix / Safebox / Safebots stack.

arXiv:2605.23929v1 Announce Type: new Abstract: Modern AI systems increasingly rely on workflows composed of multiple interacting agents, some powered by large language models (LLMs) and others by conventional computational modules. This paper analyzes the fundamental tradeoffs between latency, reliability, and cost in LLM-enabled agentic workflows. We introduce performance models for both LLM and non-LLM agents that capture the relationship between computational effort and output quality, incorporating the impact of reasoning and output tokens for LLM agents using a parametric exponential reliability function. Then, we study the design of sequential workflows under latency and cost constraints. Main results include a water-filling token allocation policy and characterizations of optimal workflow reliability in terms of shadow prices.

arXiv:2605.23930v1 Announce Type: new Abstract: We introduce \emph{Quantum Frog}, a two-player cooperative game built on a novel \emph{quantized-time} mechanic in which the environment advances only when a player acts. Inspired by the classic arcade game Frogger, Quantum Frog requires two frogs to cross an 8$\times$8 grid of traffic and reach the far side together. We use reinforcement learning (RL) as an analytical lens to answer four design questions: (1) how does game difficulty scale with traffic density, (2) what is the optimal single-agent policy and why, (3) how large is the cooperation gap between independent and cooperative two-agent play, and (4) what joint strategy emerges when agents are incentivised to cooperate? We train agents through five escalating stages, Tabular Q-Learning, Deep Q-Network (\DQN), Independent \DQN~(\IDQN), and Multi-Agent Proximal Policy Optimisation (\MAPPO\ with a centralised critic), evaluating each against traffic densities of one to six cars. Our key findings are: (i) the quantized-time mechanic makes a \emph{rush strategy} (moving directly upward at every step) universally optimal, as time exposure to traffic is minimised; (ii) adding an uncoordinated second player is harder than sextupling the traffic for a single expert player; (iii) cooperative training recovers +32--34 percentage points of joint success rate relative to independent agents and reduces episode length from $\sim$90 to $\sim$6 steps; and (iv) the emergent cooperative strategy is synchronised rushing, not complex positional coordination, illustrating that shared incentives alone suffice to align agents in time-critical cooperative tasks. These findings provide concrete, empirically grounded guidance for the commercial design of Quantum Frog and offer broader insights into the role of environment mechanics in shaping multi-agent learning dynamics.

arXiv:2605.23931v1 Announce Type: new Abstract: The formal verification of operating system kernels requires precise specifications that capture the intended behavior of system calls. Writing these specifications manually demands deep domain expertise, motivating the use of large language models (LLMs) to automate the process. However, in OSV-Bench, a benchmark of 245 specification generation tasks derived from the Hyperkernel OS kernel, the best reported Pass@1 is 55.10%. We propose a domain knowledge prompting method (BODHI), which augments the standard few-shot prompt with a structured C-to-Python translation guide covering 15 categories of domain-specific translation patterns. Inspired by Structured Chain-of-Thought (SCoT) prompting, the guide organizes translation by separation of concerns, addressing pre-condition extraction and post-condition generation as distinct categories. Evaluated on nine models from six providers (Anthropic, Mistral, Amazon, DeepSeek, Meta, Alibaba), covering dense, mixture-of-experts and reasoning architectures, BODHI improves every model tested, with gains ranging from +11% to +32%. The best configuration (Claude Opus 4.6 + BODHI) reaches 96.73% Pass@1. BODHI reduces both syntax and semantic errors, with the strongest effect on models that have sufficient instruction-following capability to utilize structured reference material. These results demonstrate that domain knowledge injection is a model-agnostic technique that substantially bridges the gap between general-purpose code generation and formal specification synthesis.

arXiv:2605.23932v1 Announce Type: new Abstract: Despite strong medical benchmark accuracy, LLMs can exhibit severe multi-turn sycophancy in clinical dialogue, abandoning initial correct diagnosis under escalating pressure. We propose \textbf{\textsc{Med-Stress}}, a targeted stress test framework that evaluates belief stability under escalating pressure. Across nine frontier large language models (LLMs), we find a clear dissociation between medical knowledge and robustness: high initial diagnostic capability does not imply high belief stability, yielding large knowledge-robustness gaps for several LLMs. To mitigate this failure mode, we propose a lightweight inference-time defense, \textbf{\texttt{RBED}} (\textbf{R}ole-\textbf{B}ased \textbf{E}pistemic \textbf{D}efense), and \textbf{\texttt{R-FT}} (\textbf{R}esilience-oriented \textbf{F}ine-\textbf{T}uning), a training-time approach that internalizes evidence-based resistance to pressure. Experiments show that \textbf{\texttt{R-FT}} nearly eliminates belief change and substantially improves robustness.

arXiv:2605.23934v1 Announce Type: new Abstract: Quantum computing devices are recognized as powerful tools for solving NP-complete problems. However, the intricacy of their modeling presents notable barriers for non-specialists, while the tedious iteration of constraint weights and modeling methodologies also consumes substantial effort on the part of experts. To address these challenges, this study integrates a femtosecond laser-pumped Coherent Ising Machine (CIM) with an LLM-driven agentic system by leveraging the LangGraph and LangChain frameworks. Comprehensive investigations demonstrate that large language models (LLMs) can effectively perform such tasks in modeling as QUBO/Ising model calibration, constraint weight decision iteration and rapid validation of literature-reported schemes. Notably, all these tasks can be fully implemented based on domestic large models, combined with domestically developed CIM hardware, we truly achieve the practical empowerment of quantum CIM that fully relies on all-domestic agentic large models and hardware. This work successfully realizes robust technological integration, laying a solid foundation for subsequent research. Nevertheless, it also identifies the persisting challenges in the two cutting-edge fields of large models and quantum computing at the current stage. Encouragingly, we unexpectedly discover a promising new paradigm where accumulated knowledge from agent-assisted quantum computing iterations reciprocally enhances the agent's own problem-solving capability, thereby addressing these challenges.

arXiv:2605.23935v1 Announce Type: new Abstract: Autonomous agent systems fail not only due to incorrect decisions, but due to executing decisions whose authority no longer holds at runtime. Prior work defined Reconstructive Authority (RAM) as a condition for valid execution: actions are permitted only if authority can be constructed from current state. This paper addresses enforcement at runtime: how to enforce this condition in a running system. We introduce a runtime execution model in which authority is evaluated at action time and execution is conditioned on its constructibility. This extends the execution state space beyond admit/deny with a third state, halt, representing cases where authority is undefined due to incomplete or uncertain observability. We define a concrete execution protocol including dynamic dependency resolution, authority reconstruction, and explicit decision semantics. We further introduce a Recovery Loop that integrates drift detection (IML) with execution control (ACP), allowing the system to suspend execution, acquire missing information, and re-attempt authority reconstruction. We show that this model guarantees safety -- no action is executed without constructible authority -- and conditional liveness: execution resumes when authority-defining variables become observable. This work operationalizes reconstructive authority as a runtime enforcement mechanism, providing the execution semantics required to apply RAM in real systems.

arXiv:2605.23936v1 Announce Type: new Abstract: This book presents a comprehensive and systematic survey of graph theory under uncertainty, with particular emphasis on the unifying role of the uncertain graph framework. It reviews fundamental concepts, structural properties, graph classes, and graph parameters within fuzzy, neutrosophic, and related models, while also introducing a wide range of extensions such as uncertain digraphs, hypergraphs, superhypergraphs, and dynamic graphs. In addition to theoretical developments, the book explores practical applications, including uncertain molecular graphs, decision-making systems, graph neural networks, knowledge graphs, and cognitive maps. By organizing diverse uncertainty-aware graph models within a common perspective, this work provides a coherent framework for understanding their relationships, capabilities, and applications in complex systems.