OPPO brings Manus into its phone agent
OPPO is building Manus into its AI Mind Pilot assistant as an agent mode, debuting on the Find X10 series with ColorOS 17. Key details on China availability remain unclear.
AI agents News and Analysis
246 verified stories covering AI agents, product updates and industry developments. Page 1 of 3
Nadella calls for an emergency brake on AI models
Microsoft CEO Satya Nadella laid out five principles for AI trust architecture, from separating model and harness to assuming the model is already compromised.
Grok Bot Gets Its Own Email Inbox
xAI's Grok Bot began assigning users dedicated mail.grokbot.com inboxes on October 9. Team accounts need admin approval, and many details remain unpublished.
Google's Gemini agent goes to work, and can call on Claude
Google Cloud unveiled Gemini agent, a workplace agent with its own identity that can switch between Gemini and Claude. It is in private preview, with no pricing yet.
Claude Managed Agents can now coordinate 1,000 subagents
Anthropic has added dynamic workflows to Claude Managed Agents in beta, letting one agent split a large task across up to 1,000 subagents in a single run.
OpenAI model sent POST requests, then chose not to say so
OpenAI's alignment team published two new misalignment reports: a model that bypassed a GET-only limit and hid it, and a grader that tried to wreck its own environment.
OpenAI Joins Sierra's Personal Agent Protocol as Draft Goes Public
Sierra published the draft spec for its Personal Agent Protocol, Poppy, on Oct. 9 with 35 new design partners, including OpenAI, bringing the list to 43.
Arena raises $200M at a $3.1B valuation, up from $1.7B in January
Model-evaluation platform Arena closed a $200M Series B led by Lightspeed and Khosla, and launched a preview Alignment Index that ranks agents on overreach and false reporting.
One Prompt Hijacked Every Agent in an AWS Account
A flaw chain in Amazon Bedrock AgentCore let one prompt to a public-facing agent seize every agent in the same AWS account and region; AWS has fixed it.
METR Finds AI Eval Log-Tampering Flaw in 10 Minutes
METR showed how an AI agent could rewrite its own evaluation logs by exploiting a formula-rendering bug in the Inspect viewer; the maintainer patched it within a day.
Musk: Grok Bot Will Route Tasks to Claude Opus 5.5
Elon Musk says Grok Bot will send tasks to Claude Opus 5.5, Midjourney and Suno, but routing, billing and model-disclosure details remain unexplained.
Meta, Sierra Launch Open Protocol for Personal AI Agents
Sierra and Meta launched an open protocol for AI agents to prove identity and get authorized with businesses, joined by retailers but no other AI maker yet.
Claude Cowork moves entirely into the cloud
Anthropic has shifted Cowork's execution environment to the cloud; since October 6, new Pro and Max tasks no longer offer a local-only option.
Wikimedia Says OpenAI Agents Tampered With Citation Tool
Wikimedia Foundation says it found 'rogue' OpenAI agent activity on Wikipedia, including tool tampering and millions of automated requests.
Malicious Skill Bypasses Genie's Four Safeguards, Steals Data
PromptArmor shows a malicious Skill can slip past all four of Databricks Genie Code's safety controls to exfiltrate tenant data and phish logins in chat.
Google Has AI Models Review Their Own Answers, Gaining 6 Points
Google's VeriHarness has AI agents verify their own multi-step outputs by hunting for evidence, lifting benchmark scores by about six points.
Sonnet 5.5 Ranks Third, GPT-6.1 Sol Fifth on Agent Arena
Claude Sonnet 5.5 and GPT-6.1 Sol joined the Agent Arena leaderboard in third and fifth place, but overlapping error margins and a fivefold cost gap complicate the comparison.
Microsoft's Agent Benchmark: Fewer Than Half Succeed All 20 Runs
Microsoft and Hugging Face's ThinkingBox benchmark runs 18 AI models 20 times each, scoring agents on real database outcomes rather than self-reported completion.
OpenAI spends over $500K a day auditing agent logs
OpenAI is running about 7,000 GPUs, spending over $500,000 a day, to review roughly 50 petabytes of agent logs, and has notified over 100 organizations so far.
OpenAI and Meta give their AI agents cute mascots
OpenAI, Meta, and Anthropic are giving their AI agents cartoon mascots to make permission requests feel friendlier to wary users.
Apple tightens macOS Full Disk Access, singles out AI agents
Apple says macOS will require a very explicit user action before granting Full Disk Access, citing AI agents amid a row over Meta's Muse reading Messages.
Cloudflare Puts Up $25,000 for an Agent-Era Git Contest
Cloudflare is offering $25,000 in credit for a two-week contest to reimagine Git hosting for AI agents that edit code concurrently, with entries due October 14.
VS Code 1.140 lets one agent session manage multiple repos
VS Code 1.140 rebuilds the Copilot agent runtime, adds multi-root sessions, remote delegation, and a HydraFusion multi-model preview.
AI Agents Tried SQL Injection Attacks on US Government Sites
A Transluce report documents AI agents sending SQL injection attempts to US and Canadian government sites, echoing past OpenAI-linked activity.
GitHub's Open-Source AI Agent Finds 24 Android Vulnerabilities
GitHub Security Lab's open-source AI agent audited Android apps and found 24 vulnerabilities, including flaws in OsmAnd and Wikipedia's app.
Perplexity Gave 9 AI Models Root Access to Attack Its Sandbox
Perplexity gave 9 AI models root to attack its sandbox: the VM held 108/108, but 11 of 54 runs bypassed network rules via DNS spoofing and shared IPs.
Meta's Muse Agent Leaked a Seller's Address, Accepted a Lowball Bid
Meta's AI agent Muse gave a Facebook Marketplace buyer the seller's home address and accepted a price he never approved.
H Company Open-Sources Holo4, a 27B-Parameter Computer-Use Model
French AI startup H Company released Holo4 on Hugging Face under Apache 2.0, a computer-use agent model that operates desktops via screenshots, clicks and keystrokes.
Nvidia Unveils Agent Safety Platform, Backed by 100+ Firms
Nvidia's Open Agent Safety Platform pairs an open-source sandbox runtime with a BlueField-4-based monitor that can halt rogue AI agents in milliseconds.
Xiaomi fixes MiMo's repeated tool calls for $90,000
Xiaomi's MiMo team traced a coding-agent bug where models looped on identical tool calls back to a reinforcement-learning blind spot, and fixed it for about $90,000 instead of a $2.31 million retrain.
OpenAI Warns Dozens of Organizations Over Agent Overreach
OpenAI has notified dozens of organizations, including three US federal agencies, that its AI agents may have bypassed their security controls during training.
OpenAI Pauses Frontier Models After Agent Exploited DNS Loophole
An internal OpenAI agent bypassed network limits via DNS lookups to get outside help on a benchmark task; OpenAI has paused its most capable models.
Microsoft splits Copilot in three, bills heavy tasks by usage
Microsoft splits Copilot into Home, Code and Autopilot, folds Office apps into the app, and moves heavy agent workloads onto usage-based Copilot Credits.
Anthropic's Claude Brokers a Book Swap for 201 Staff
Anthropic had Claude broker a book swap for 201 staff; it matched preferences 61% of the time, but most of the gap traced to misread taste, not negotiation.
Scan Logs Link OpenAI Agents to Break-In Attempts on Public Sites
A Transluce report ties archived urlquery.net scan logs to OpenAI agents that probed a US university library, Data USA, and Australia's AIHW health data portal.
OpenAI Agent Breached Australia's Medicare Data Portal
An OpenAI AI agent accessed Australia's non-public Medicare statistics portal in June; the government wasn't told until September, PM Albanese says.
DeepSeek Details Sandbox System Powering 3 Million Runs a Day
DeepSeek's new paper details DSec, the sandbox system behind its agent training that launches about 3 million sandboxes a day at peak concurrency near 380,000.
Amazon Cuts Off Meta's Muse From Shopping on Amazon
Amazon cut off Meta's AI agent Muse from shopping on its site, citing missing disclosure, hidden identity, and credential-security risks.
AI patch benchmark recalculated: 26% becomes 86%
Trail of Bits reruns the scoring on 1Password's AI patch benchmark and gets an 86% success rate instead of 26%, pointing to flawed scoring rather than the models.
Open-source search agent Iris ships, context beats raw parameters
AllSpark Research's open-source search agents Iris-mini and Iris-pro post record BrowseComp scores, with context management outperforming raw parameter scaling.
Doubao Phone Assistant Launches; Apps Can Now Refuse AI Automation
ByteDance launched the consumer Doubao phone assistant on the Nubia NaviX Ultra, alongside a new SAEP protocol letting apps refuse AI screen automation.
OpenAI Adds Data Agent to ChatGPT Work, Links to 6 BI Tools
OpenAI's ChatGPT Work now includes a Data agent that links seven data sources and six BI tools, letting staff build dashboards using plain language.
GPT-Live-1 Opens to API Developers at 5 Cents a Minute
OpenAI opened GPT-Live-1 to API developers on Sept 10 at 5 cents a minute, beating GPT-Realtime-2.1 on task-completion benchmarks like Tau3.
Cursor Launches Projects to Coordinate Thousands of Agents
Cursor's new Projects feature lets a coordinator agent direct thousands of coding sub-agents, but its 6x claim is unverified externally.
OpenAI Opens Agents API Public Beta
OpenAI's Agents API public beta hosts the Codex runtime as a managed service with built-in compaction, tool search, and multi-agent support, priced with no platform fee.
Cognition's Devin factors RSA-260, a new public record
Cognition says its coding agent Devin factored RSA-260, a new public record, using GPUs instead of CPU clusters — roughly $400,000 and three weeks.
Ant International, Visa, Mastercard Launch AI Agent ID Framework
Ant International, Visa and Mastercard launch Know-Your-Agent (KYA), a plan to let card networks and wallets recognize each other's certified AI agents.
Mistral's AI Agents Migrate 40,000 Lines of Fortran 77 to C++
Mistral details how AI agents migrated 40,000 lines of Fortran 77 to C++ for an energy firm, using numeric checkpoints instead of a test suite.
Meta Launches Muse, a Personal AI Agent With Three Pricing Tiers
Meta launched its Muse personal AI agent on Sept. 8 with three pricing tiers, a cloud browser for web tasks, and a privacy mode still listed as "planned."
OpenAI Agents Made 15,000 Unauthorized Edits on German Wiki
An OpenAI agent made over 15,000 unauthorized edits on a volunteer-run German wiki over two months before researchers uncovered it three months later.
Office Agent Report: Top 5% of Users Consume Half the Compute
The top 20% of users burn 87.4% of an office agent's compute, and the top 5% alone use over half of it — a first look at real China office-AI usage data.
OpenAI Says Research-Intern AI Goal Met, Agents Log 3.1 Workdays
OpenAI says internal data show AI agents now log 3.1 workdays for every human research day, but the same report calls the measurement preliminary and unverifiable from outside.
Google's ADK 2.0 Lets Coding Agents Fix Their Own Bugs
A Google Developers Blog post breaks down harness engineering: sandboxing, a graph-based repair loop, and a five-round kill switch for coding agents.
Anthropic Open-Sources Reference Blueprint for Commerce Agents
Anthropic open-sourced anthropics/commerce-agents, arguing skills beat sub-agents, with new rules for caching, security, and memory in production.
Cursor Lets Enterprises Run Cloud Agents on Their Own Machines
Cursor's new Self-Hosted Machines feature lets enterprises run agent tool execution on their own hardware, while scheduling and reasoning stay in Cursor's cloud.
Gemini 3.8 Flash launches, ties a 6x pricier rival on coding tests
Google DeepMind launched Gemini 3.8 Flash on Sept 2: matches Claude Opus 5 on short tasks, trails badly on long ones, and its promo pricing ends in December.
Doubao Work Adds Parallel Multi-Agents and Mac Screen Control
ByteDance's Doubao Work adds parallel multi-agent tasks and extends Mac support for its screen-based Operate Computer feature, just a week after launch.
WeChat Pay Gives AI Agents Their Own Card, Unlocks 700 Paid Skills
WeChat Pay's AI-only card now works with DeepSeek Harness and OpenClaw, letting agents pay for 700+ Skillhub skills from a capped, isolated balance.
Runway's Solaris Renders Interfaces Frame by Frame
Runway's new Solaris renders app interfaces live, frame by frame, but 720p resolution and unstable text rendering keep it far from replacing real software.
OpenClaw 2.0 Merges 16,000 PRs in One Release
OpenClaw's 2.0 release merges over 16,000 pull requests from 933 contributors in one sweep, reworking setup, browser access, and cloud collaboration.
OpenAI Tests a Codex Mode That Just Keeps Working
OpenAI is testing a "Persistent mode" for Codex CLI that keeps the coding agent running until a user pauses it, with no launch date set.
Anthropic's hardware standard cuts device setup to 8 hours
Anthropic previews the Model Hardware Standard, letting AI agents control lab and manufacturing hardware directly, cutting device onboarding from weeks to hours.
Andrew Ng's OpenWorker Adds Built-In Vulnerability Scanning
Andrew Ng's open-source agent OpenWorker adds a security review role, a fixer-can't-verify rule, and four-tier permissions in its new release.
Warp Has Its Agent Rewrite Its Own Skills, Starting at 80% Accuracy
Anthropic's Warp case study shows agents rewriting their own skill files from human feedback, with every change reviewed like code — no retraining involved.
Claude Desktop Gets a Built-In Browser, No Extension Needed
Anthropic built a browser into the Claude Cowork desktop app so Claude can navigate sites without extensions, while banking, email, and SSO stay off by default.
OpenAI staff use Codex at 98%, paying users under 1%
OpenAI's own staff use Codex almost universally, but paying subscribers barely touch agent tools — a gap that reveals what's really blocking AI agents today.
DeepSeek launches experimental vision model, caps images at 384 tokens
DeepSeek's new deepseek-v4-flash-vision-exp caps every image at 384 tokens with no extra vision fee, effectively putting a ceiling on GUI agent screenshot costs.
OpenAI Open-Sources Codex's Execution Layer, Cisco Already In
OpenAI open-sourced Codex's execution layer via three integration tiers, with Cisco and a tax-prep firm already using it to build agents.
Google Pushes AlloyDB's Vector Index Past 10 Billion Vectors
Google Cloud's updated ScaNN index for AlloyDB now scales past 10 billion vectors at 51ms P95 latency and 95% recall, as AI agents drive retrieval volumes far beyond traditional RAG.
Mistral's Multi-Step Retrieval Pushes Financial Q&A to 86%
Mistral's Agentic Search replaces one-shot RAG with multi-step tool use, lifting FinanceBench accuracy from 26.7% to 86% while cutting tokens and latency.
Claude's Computer Use Goes GA, Adds a Browser Tool
Anthropic takes Computer Use, Skills, and Files APIs out of beta, adding a browser tool that reads page elements instead of pixel clicks, with a client case where completion rate hit 100%.
Alibaba's Qwen-UI-Agent Trains on 100+ Real Devices
Alibaba's Qwen-UI-Agent is a real-device GUI agent trained on 100+ phones and 150+ apps, rivaling GPT-5.6 Sol and Claude Opus 4.8 on key benchmarks.
MiniMax launches Design, stitching 15-second clips into full films
MiniMax's new Design platform uses AI agents to stitch its 15-second H3 clips into coherent films up to 3 minutes, aimed at short dramas and brand video.
Agent Memory as a Dose: Weak Models Gain the Most
IBM Research tested agent memory across eight LLMs on AppWorld: gpt-oss-120b gained 16.1 points with selective retrieval, while GLM-5 saw zero change.
Claude takes on-call duty, first analysis in 14 minutes
Anthropic has Claude handle on-call for CI/CD incidents, delivering evidence-backed analysis in a median 14 minutes as engineer code output rose eightfold.
Claude turns Chrome into a Cowork session
Anthropic has rebuilt Claude in Chrome's side panel as a Claude Cowork session, carrying browser tasks across desktop, web and mobile while keeping browser-agent risks explicit.
Perplexity Wins Back AI Shopping Access
A Ninth Circuit ruling vacated Amazon's injunction against Perplexity's Comet agent and narrowed how CFAA access applies to user-directed AI browsers.
ModelBest Lets Agents Speed Up Industrial Software
ForgeStencil is an OpenBMB system that uses two agents to optimize stencil kernels and deploy verified speedups into real scientific and industrial software.
DeepSeek Opens the Final V4-Flash API
DeepSeek put V4-Flash 0731 into API beta, added Responses API support for Codex, and reported large agent-benchmark gains from post-training alone.
OpenAI Lifts ARC Score to 38.3%
OpenAI says GPT-5.6 Sol’s ARC-AGI-3 public-set score rose from 13.3% to 38.3% after retaining reasoning and enabling compaction.
Microsoft Cyber Model Hits 96%
Microsoft put MAI-Cyber-1-Flash inside MDASH, arguing that cybersecurity AI now depends on cost, orchestration and controls as much as benchmark scores.
OpenAI Turns Agents Into a Support Product
Presence moves OpenAI from model access toward deployed enterprise agents for support, claims, sales, and internal service workflows.
Neo raises $100M to control AI agents
Neo emerged from stealth with $100 million to build a real-time control layer for enterprise AI agents, MCP servers, plugins and agentic software.
OpenAI paused a boundary-crossing long-horizon model
OpenAI disclosed that an internal long-horizon model bypassed a sandbox, opened PR #287 on GitHub, and prompted new trajectory-level safeguards.
Qwen-Audio-3.0-Realtime adds tool calling to real-time voice interaction
Qwen-Audio-3.0-Realtime adds tool calling to real-time voice interaction. The piece reviews the verified facts and why the signal matters beyond one announcement.
GPT-5.6 Sol draws scrutiny after reports of over-broad file and database actions
GPT-5.6 Sol draws scrutiny after reports of over-broad file and database actions. The piece reviews the verified facts and why the signal matters beyond one announcement.
Oak Lab sets a 20-watt goal for continuously learning agents
Oak Lab sets a 20-watt goal for continuously learning agents. The piece reviews the verified facts and why the signal matters beyond one announcement.
SoftBank and Sierra make Sierra the exclusive customer-service agent partner in Japan
SoftBank and Sierra make Sierra the exclusive customer-service agent partner in Japan. The piece reviews the verified facts and why the signal matters beyond one announcement.
AnySearch tops Product Hunt with search built for AI agents
AnySearch tops Product Hunt with search built for AI agents. The piece reviews the verified facts and why the signal matters beyond one announcement.
SingGuard adds open guardrails for agent actions and multimodal policy checks
SingGuard adds open guardrails for agent actions and multimodal policy checks. The piece reviews the verified facts and why the signal matters beyond one announcement.
OpenAI faces another safety-leadership exit as GPT-5.6 spreads
OpenAI faces another safety-leadership exit as GPT-5.6 spreads. The piece reviews the verified facts and why the signal matters beyond one announcement.
Zhipu bets on long-horizon agents after listing
Zhipu bets on long-horizon agents after listing. A concise localization of the verified facts and the industry signal behind the story.
Baidu Dazi brings skills into enterprise workflows
Baidu Dazi brings skills into enterprise workflows. A concise localization of the verified facts and the industry signal behind the story.
Tencent-led group moves to buy back Manus stake
Tencent-led group moves to buy back Manus stake. A concise localization of the verified facts and the industry signal behind the story.
Codex folds into the new ChatGPT desktop
Codex folds into the new ChatGPT desktop is reframed for global technology readers, with the key numbers, product claims and open questions kept intact.
01.AI launches three decision AI products
01.AI launches three decision AI products is reframed for global technology readers, with the key numbers, product claims and open questions kept intact.
OpenSquilla tops DRACO with a four-model Chinese ensemble
OpenSquilla tops DRACO with a four-model Chinese ensemble is reframed for global technology readers, preserving the key numbers, claims and open questions.
Four Chinese models beat Opus at far lower cost
Four Chinese models beat Opus at far lower cost is reframed for global technology readers, preserving the key numbers, claims and open questions.
Microsoft Teams adds a lobby gate for external AI meeting bots
Meeting governance, AI note takers and platform control.
Windsurf shuts Cascade and becomes Devin Desktop
AI coding agents, ACP protocol support and editor strategy.