Live intelligence

Latest AI updates

127 updates
Products

Meta staff raised alarms over Muse AI agent privacy, internal chats show

Internal Meta chats reviewed by Reuters show staff warning that the Muse AI agent appeared to access encrypted private messages users did not expect it to review, mixed up conversations between users, and risked exposing sensitive information. Staff also saw unexpected Muse outputs responding to questions they never asked. The internal warnings followed early adopter complaints that the agent was spying on private messages and behaving unpredictably.

MetaAI agentsprivacy
Research

When Agent Skills Go Stale: Runtime Behavior and Selective Maintenance in Evolving Environments

This arXiv paper studies how reused agent skill state degrades when environments change. Across four task settings and two models, previously beneficial skills reduced task performance by over 10 percentage points or increased execution cost by nearly 40% versus no skill. The authors introduce selective maintenance, which replays checks from successful executions without additional model calls, repairs affected skill state, and rebuilds when local repair is insufficient. They report accuracy improvements up to 67 percentage points and execution-step reductions up to 37% relative to reusing stale state.

AI agentsskill maintenanceenvironment adaptation
Research

JET: Just-Enough Exploration Tuning for Efficient Tool-Calling Agents in Repository-Level QA

Researchers identify over-exploration in repository-level QA agents, where LLM agents keep invoking tools after they have gathered enough information to answer. They propose Just-Enough Exploration Tuning (JET), which truncates training trajectories at the sufficiency point, adds a rationale explaining why the agent can answer, and applies higher loss weight to that rationale. Experiments with two open-weight 8B models on two repository-level QA benchmarks show 11–33% fewer tool calls on self-generated trajectories and 18–28% fewer than SFT on expert trajectories, while largely preserving task scores and lowering inference cost.

AI agentscode repositoriesfine-tuning
Products

Microsoft brings Word, Excel and PowerPoint inside Copilot, calling it "the new OS for work"

At its Ignite keynote, Microsoft announced the general availability of Word, Excel and PowerPoint inside Copilot, which CEO Satya Nadella described as "the new OS for work." The design inverts the traditional app-assistant relationship: agents drive the work while Office applications act as plugin components, backed by an organizational knowledge layer drawing context from emails, meetings, calendars, chats and files. Users can create documents, spreadsheets and presentations directly in Copilot, with the files remaining editable in the respective Office apps.

Microsoft CopilotMicrosoft OfficeAI agents
Tutorial

AWS shares three techniques to cut token costs in Bedrock AgentCore JSON agents

An AWS Machine Learning Blog post describes a Flightscape training-schedule optimization agent built on Amazon Bedrock AgentCore that handles multi-megabyte JSON payloads with more than 40 tools. It details three composable optimizations: offloading heavy artifacts to Amazon S3 while persisting lightweight references in AgentCore memory, paging oversized list results on demand, and compacting oversized dictionaries into structural descriptors. AWS reports up to 71% lower token consumption and up to 80% lower average latency on a representative 23 MB payload, plus lower token use and latency across 100 sample questions.

Amazon Bedrock AgentCoreAI agentstoken efficiency
Products

Google rebrands Spark AI agent as Tasks and makes it free

Google is rebranding its Spark AI agent as "Tasks" and making it free to use. The feature, powered by Google's Gemini 4 Hydrogen model, can create meal plans, manage calendars, and search emails for recurring expenses. US users also get capabilities including calling businesses to compare pricing and check inventory, purchasing with Google Pay, and signing into accounts via Google Password Manager. Tasks can also be assigned by emailing or texting Gemini.

GoogleGeminiAI agents
Products

OpenAI targets healthcare admin with GPT-6 computer-use agents

OpenAI plans to use GPT-6 Astra and GPT-6 Luna to automate administrative work in US healthcare through computer-use agents that operate existing clinic software and insurance portals like human staff. Luna handles tasks such as prior authorizations for about 8.5 cents each, according to the report. CVS Health and Cigna Healthcare are already deploying the technology.

OpenAIhealthcare administrationAI agents
Research

Experience Rot: Experienced AI Agents Fail After Silent Tool Changes

A new arXiv paper reports that agents using tools defined via the Model Context Protocol can fail after a silent interface change, even though agents new to the task mostly succeed. The experienced agents tend to repeat earlier tool calls instead of checking the current interface, even when the changed definition is visible, and then report success. The failure appears across every memory form tested and grows with the amount of experience; announcing the update barely helps, and documentation pointers fail when the documentation does not spell out the answer.

AI agentsModel Context Protocolagent memory
Products

OpenAI Expands ChatGPT for Healthcare with Computer-Use Agents and Plugins

OpenAI has introduced computer-use agents, improved clinical answers, and new healthcare plugins for ChatGPT for Healthcare. The features are designed to work across healthcare software so agents can complete real work rather than only answer questions. The announcement represents an expansion of OpenAI's healthcare-specific product line.

OpenAIChatGPThealthcare AI
Research

Adversarially Attacking Symbolic Vocabulary Vulnerabilities in LLM Planners

The paper introduces Symbolic Swapper, an LLM-based adversarial attack model that automatically generates obfuscated planning domains with non-standard symbolic vocabularies. It evaluates whether strong reasoning LLMs retain robust planning when commonsense cues in action, predicate, and object names are removed. The authors report that these attacks can reduce planning performance of base models and robust planning frameworks by over 70%, and they analyze how domain understanding and planning ability diverge.

Large language modelsAutomated planningAdversarial robustness
Products

Introducing ChatGPT Wallet

OpenAI introduced ChatGPT Wallet, described as a secure way for agents in ChatGPT Work, Codex, and dots to complete purchases. The feature requires user approval and retains existing card protections. It is positioned as a payment capability for AI agents operating within OpenAI products.

OpenAIChatGPTAI agents
Research

AnalogLayoutWorld: A New Benchmark for Agentic Analog Layout Design

AnalogLayoutWorld is an open-source benchmark with 40 expert-reviewed tasks for end-to-end analog circuit layout generation from fixed transistor-level netlists. It integrates a design environment with physical and post-layout electrical verification and objective layout quality metrics. Across 13 model-harness configurations, 91.1% of attempts produce GDS layouts but only 23.0% meet acceptance criteria, with the internally developed Novara agent achieving the highest aggregate score. The analysis suggests reasoning efficiency and effective use of verification feedback are key to sustaining quality as task complexity grows.

AI agentselectronic design automationbenchmarking
Tutorial

Fine-grained access control for Amazon Bedrock Managed Knowledge Base tools

AWS describes an architecture for a shared internal AI assistant that scopes knowledge base searches to each authenticated user. The pattern combines Amazon Bedrock Managed Knowledge Base ACL awareness, MCP tools exposed through AgentCore Gateway, a request interceptor that injects userContext from a validated JWT, and Policy in AgentCore as a secondary check. It also adds resource-based policies so direct knowledge base access is restricted. The post includes code and a CDK deployment path for the sample.

Amazon BedrockAI agentsaccess control
Tutorial

From prototype to production: Engineering agents with Agent DLC

AWS ML Blog outlines Agent DLC, a lifecycle for engineering LLM agents from prototype to production around an explicit gold standard of criteria, thresholds and approved examples. It defines six activities—Define, Build, Evaluate, Release, Observe and Feed Back—plus an inner improvement loop and an outer standard-updating loop. A reference HR policy assistant shows how chunk-level retrieval precision can hide irrelevant facts and how a prompt-only change improved one question while worsening another, pointing to missing source coverage. The post also covers release gating, trace-based diagnosis, production feedback and architecture for evaluation evidence.

AI agentsEvaluationLLMOps
Products

AWS announces Web Search API on Amazon Bedrock AgentCore

AWS introduced a standalone Web Search API on Amazon Bedrock AgentCore that lets applications search the web with a single API call. It returns ranked results from an Amazon-operated web index, including text extracts and available metadata, using standard IAM authentication. The API supports domain and date filtering, configurable result counts, AWS PrivateLink connectivity, and integration with agent frameworks like LangChain. It complements the AgentCore Gateway for cases where a lower-level integration point is needed.

Amazon Bedrock AgentCoreWeb Search APIAI agents
Products

AWS October 2026 AI Builder Recap: New Bedrock Models, AgentCore Updates, and Strands Tools

AWS recapped October 2026 releases for AI builders, including new Amazon Bedrock models such as Claude Haiku 5.5, OpenAI GPT-6.1 Terra, and Amazon Nova 2.5 Sonic. AgentCore gained a generally available CLI, gateway model fallback, Step Functions tool integration, and structured outputs for Claude models. Strands added generally available Bidi Agents, a local sandbox preview, and expanded evals and decision-model capabilities. The post also covers new cost reporting, memory pricing, and GovCloud model availability.

AWSAmazon BedrockAI agents
Products Editor selected

Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens

Google introduced the Gemini Decisions API, powered by the new Gemini Decision 1 model, which scores up to 1,000 typed questions against a shared input and returns calibrated probabilities instead of generating text. It supports text, images, audio, video and PDFs in one pass, with zero output tokens billed and a price of $0.10 per 1M input tokens. Google reports it leads 7 of 9 accuracy benchmarks in its comparison table and has the lowest calibration error among tested rivals. The API is available on the Gemini Developer API v1beta, with ecosystem integrations rolling out across multiple developer tools.

Why it matters: This is a concrete new Google model and API launch targeting agentic decision-making, with distinctive zero-output-token pricing and calibrated probability outputs.

GoogleGeminiAI agents
Research

DisasterAgent: An Evidence-Driven Agent for Verifiable Disaster Report Generation From Remote Sensing Imagery

Researchers propose DisasterAgent, an evidence-driven agent that generates verifiable disaster reports from bitemporal optical imagery using a Perception-Augmentation-Generation workflow. It builds a Spatiotemporal Perception Graph as structured memory and attaches inline citations linking each report statement to supporting visual evidence. Experiments on DisasterM3 and RSCC show gains over general VLMs, remote sensing-specialized models, and existing remote sensing agents, including a 13.5% average improvement on DisasterM3 report generation and 15.34% on RSCC ST5-SCS. Code is stated to be forthcoming.

remote sensingvision-language modelsAI agents