AI News Today — November 8, 2026: Amazon Texts You After Blocking Muse, Claude Code Backgrounds Long Jobs, and Google's Agent Kit Goes Stable

The big picture

This was a quiet weekend for launches — no major new model landed in the last 24 hours — and that makes the real story easier to see: the agent wars have turned territorial. Amazon, which blocked Meta's Muse agent from querying its store in September, just launched its own text-message shopping agent, in the same style Muse uses, timed for the holiday season. While the consumer agents fight over whose storefront they may enter, the developer infrastructure underneath them matured in public: Anthropic's Claude Code now backgrounds long-running tool jobs instead of blocking on them, and Google's Agent Development Kit graduated agent-to-agent communication to stable. And the security bill for connecting agents to everything arrived twice in one day — OpenAI formally hired the security firm that breached its own systems to test its third-party connections, while researchers found hundreds of AI hacking tools sitting exposed on the open web. Agents are getting more connected, more capable, and more worth locking down — all at the same time.

What happened today

Amazon lets you text Alexa to shop — after blocking Muse from its store

What happened: Amazon announced on November 6 that select U.S. customers can now text its Alexa for Shopping agent to search for products, compare prices, add items to a cart, and complete purchases in natural language. Invited customers get a text asking them to enable the feature; everyone else can join a waitlist, and Amazon says it will roll out more broadly — to millions more customers, per its announcement coverage — in the coming weeks, just before the holiday shopping season. The text agent builds on Alexa's recent shopping upgrades: it can manage lists, track prices, search across the web, check order status, and search for similar listings from a photo of an item or an outfit. The context is the story: TechCrunch reports Amazon blocked Meta's Muse from querying its site in September, and TechCrunch notes the new feature works "much like you can with AI agents like Muse or Instinct." Amazon's own announcement does not mention Muse. Amazon says Alexa for Shopping was used by 400 million+ customers in the last 12 months, per its third-quarter release.

Who it matters to: Anyone building — or using — a shopping agent. The largest storefront in U.S. e-commerce has now stated its position by action: outside agents do not get to shop here, but Amazon's own agent gets the most intimate channel there is, your text thread. DoorDash reached a similar conclusion last month, per TechCrunch's reporting: third-party agents had placed only hundreds of orders on its platform, so it is concentrating on its own text-based agent instead. The pattern for agent builders is blunt — distribution through someone else's storefront is a privilege that can be revoked, not a right.

What you can do: If you're a U.S. Amazon customer, check whether you've been invited, or join the waitlist, and test the agent the way you'd test any shopping agent: start with price tracking and list management, and review the cart yourself before any purchase. If you sell on Amazon, note that product discovery is moving into a text thread you don't control — your listings' titles, images, and price accuracy are now what an agent reads on a customer's behalf.

Learn more: TechCrunch on Amazon's text-to-shop Alexa launch

Muse AI: no new launch — Spaces are now available in every Muse market

What happened: Our daily priority check found no new Muse AI announcement on November 7–8. What changed is confirmation and reach: coverage published November 8 confirms Spaces — the feature covered in yesterday's briefing, where Muse builds a tracker, planner, game, or dashboard from a chat description and keeps it updated from your connected apps — is available in all markets where Muse is offered, with Meta dating availability from November 7. The Wonder-trained specialized models with Active Memory and Active Focus continue rolling out to all Muse and Meta AI users on the same terms as announced November 6: Meta's performance figures (76% to 90% correct use of critical context; Personal Superintelligence Index 38% to 47%) remain Meta's own internal evaluations, with no model names, sizes, baselines, or methodology published, and no third-party validation yet. The other Muse development this weekend is the one in the story above: the agent Amazon blocked in September is now the comparison point for Amazon's own launch — Muse was the example TechCrunch reached for to explain what texting Alexa is like.

Who it matters to: Muse users outside the first wave, first of all: Spaces and the model upgrade are no longer a U.S.-only story — they are live wherever Muse is offered (the U.S., Canada, Japan, Singapore, and Mexico). It also matters as a market signal. In six weeks Muse went from launch to 10 million users in the U.S. and Canada, per Meta's October earnings call — and to being the agent a trillion-dollar retailer both blocks and imitates in the same quarter.

What you can do: If you're in a Muse market and haven't tried Spaces yet, build one this weekend for one recurring job — a holiday gift tracker is Meta's own example and the season makes it a real test — connect only the apps that Space needs, and check in a few days whether its background updates stayed accurate. And if you shop with agents, run the comparison Amazon just set up: ask Muse and Alexa for the same product and compare what each can actually see and do.

Learn more: Storyboard18 on the Spaces launch and availability

Claude Code 2.1.319: long MCP jobs go to the background, and unattended runs can deny instead of asking

What happened: Anthropic released Claude Code 2.1.319 on November 6. The headline change: when an MCP server runs a tool call as a task, Claude Code now moves it to the background right away and tells Claude the result when it finishes, instead of blocking the session on a job that might run for twenty minutes. An environment variable, CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND, restores the old wait-inside-the-call behavior. MCP tasks leave their feature-flag gating behind in this release, so Bedrock, Vertex, Foundry, and LLM-gateway setups get them too. Second, a new setting, permissions.denyInsteadOfPrompting, denies any tool call that would normally stop and wait for a permission prompt, and lets the turn continue — built for terminal sessions, subagents, and teammates running unattended, where nobody is there to click "Allow." Third, auto mode tightened behind LLM gateways: calls the server can't check are now asked about or denied, with no fallback to the built-in classifier. The release also adds a /claude-api latency-optimize command and fixes several permission-check bypasses, including a crafted environment prefix in Bash and a sandbox escape during worktree creation, per the changelog summaries.

Who it matters to: Anyone who runs coding agents unattended — overnight jobs, CI pipelines, multi-agent teams — which is increasingly how Claude Code is used. Background MCP tasks remove a real friction: one slow report export or build no longer freezes the whole session. The deny-instead-of-prompting setting is the more consequential change, because it makes the safe failure mode the automatic one: an unattended agent that hits something it wasn't pre-approved to do now skips it and keeps working, instead of either hanging forever or, worse, being run with prompts disabled entirely to avoid the hang.

What you can do: Update to 2.1.319, then do two things. For any unattended or scheduled Claude Code runs, turn on permissions.denyInsteadOfPrompting and pre-approve exactly the tools those runs need — anything else gets denied and the run continues, which is the behavior you want at 3 a.m. And if you use MCP servers with long-running tools, test one task-type call and confirm it backgrounds and reports back; if a workflow genuinely needs the result inline, set the disable variable for that workflow only.

Learn more: Releasebot's summary of the Claude Code 2.1.319 changelog

Google ADK 2.14: agent-to-agent communication and execution environments graduate to stable

What happened: Google published the Agent Development Kit (ADK) Python v2.14.0 release tag on November 7 (the release itself is dated November 4). Its headline is graduation: execution environments, A2A (agent-to-agent) integration, evaluation, and core runtime modules move from experimental to stable — 26 auth, environment, app, CLI, and cache modules in all, plus 16 toolset and integration features. The stable execution environments include local, Daytona, and E2B backends. New in the release: a TeamsRunner that connects ADK agents to Microsoft Teams chats and channels, a MultiRunConfig that runs each evaluation case multiple times concurrently, a MongoDB memory service with vector-search retrieval and TTL expiration, and an adk generate_agent_card command that builds an A2A agent card from the CLI. Google claims a 30-node workflow running 200 turns is 23% faster with 9% lower peak memory in this release — and publishes no baseline hardware, comparison version, or second benchmark to check that claim against.

Who it matters to: Teams building agents on Google's stack, and anyone watching agent interoperability stop being a slide and start being a stable API. A2A graduating matters because cross-agent work — one agent handing a task to another agent built by a different team — needs a protocol both sides can depend on not changing next month. Stable status is Google saying these surfaces are now a baseline you can build production plans on. It is not, by itself, a guarantee that any given application is production-ready; your own security and reliability checks still decide that.

What you can do: If you build with ADK, pin 2.14.0 and move any experimental-environment or A2A code onto the stable surfaces while the migration is small. Generate an agent card for one existing agent with the new CLI command — it's the cheapest way to see what your agent advertises to other agents. And treat the 23% speed claim as untested until you've benchmarked your own workflow: Google didn't publish the setup you'd need to reproduce it.

Learn more: DEV's breakdown of the ADK v2.14.0 release notes

OpenAI hires Hacktron — the firm that breached it — to test its third-party connections

What happened: OpenAI announced a partnership with security firm Hacktron on November 6 to find vulnerabilities in its systems, concentrating on connections with third-party services. OpenAI engineers and Hacktron researchers will jointly select high-priority areas to test, and Hacktron gets expanded access to frontier models through Daybreak Red, OpenAI's program for authorized vulnerability research, with identity verification, monitoring, and approved-use controls. The relationship began in OpenAI's bug bounty program, and Hacktron's own account shows why OpenAI takes it seriously: on July 25, 2026, its researchers chained two critical vulnerabilities — remote code execution in the community forum's image-processing path, plus an OpenAI single sign-on issue — to compromise multiple OpenAI employees' ChatGPT accounts, reaching toward internal repositories and a connected GitHub integration, before stopping at a harmless proof pull request. Hacktron says OpenAI confirmed a fix roughly 14 hours after the report; OpenAI paid a $6,500 bounty on September 1. Coverage of the announcement also reports OpenAI has raised its bug bounty rewards.

Who it matters to: Anyone connecting agents and AI tools to other services — which, after this week's Muse Spaces, Gemini Tasks, and Alexa launches, is the entire direction of the industry. Hacktron's July chain is the template for how these systems actually fall: not through the model, but through the seams — a forum's image library, a sign-on flow, a connected integration — where access granted to one system silently carries into the next. OpenAI handing its own attack surface to the firm that proved that point, with cyber-capable models included, is the defender version of the same insight.

What you can do: Map your own seams before someone else does. List every third-party connection your AI tools hold — sign-on providers, GitHub and Drive integrations, MCP servers, connected apps in your agents — and remove the ones you no longer use; each dormant connection is a path like the one Hacktron chained. Where a connection stays, scope it to the least access that works, and make sure a compromise of the least important connected service cannot sign into the most important one.

Learn more: Unite.AI on the OpenAI–Hacktron partnership and the July breach chain

Ai2's ToyBlocks: an open robot world where failures can be traced, not just counted

What happened: The Allen Institute for AI, Ai2, released ToyBlocks on November 6 — the first environment in its FoundationWorlds framework, and a genuinely open one: simulation, code (Apache-2.0), dataset, and two trained models, all public. The world is deliberately small — a simulated robot arm and labeled blocks on a tabletop, matching Ai2's real Seattle lab — so that experiments are repeatable and failures diagnosable. The initial dataset holds 85,336 long-horizon demonstrations covering nearly 1,600 hours of robot execution and 144 million frames, generated automatically by planners at up to 24 times real-time on one GPU. The released policy is two-level: Molmo2-4B-tb plans from a camera image and a text goal, and ManiFlow-tb turns the plan into actions. The results are honestly mixed, and Ai2 says so: the hierarchical system succeeded in 55.5% of simulated episodes and 50.0% of physical episodes on a real Franka arm, trained entirely in simulation — while ManiFlow, π0.5, and MolmoAct 2 all showed poor generalization despite large-scale pretraining. Ai2's conclusion: action-model generalization remains far from solved.

Who it matters to: Robotics researchers get a benchmark where a failure can be decomposed — swap in an oracle planner or an oracle actor and find out whether the decision or the movement was wrong — instead of another leaderboard number. Everyone else should take the generalization finding seriously: models that perform inside their training setup and fall apart outside it is the same failure mode appearing in agents generally, here measured cleanly because the world is blocks, not the open web. A second release with more skills and integrated reinforcement learning is planned, but is not in this package.

What you can do: If you work in robotics or sim-to-real research, ToyBlocks is free to download and run today — start with the easy/medium/hard dataset splits and the withheld-task evaluations, and compare your policy's planning errors against its execution errors using the oracle-swap setup. If you don't, the transferable lesson is the audit method: when your own agent fails, separate "chose the wrong action" from "executed the right action badly" before you retrain anything.

Learn more: Unite.AI on the ToyBlocks release, data, and results

Gemini Spark: no new Spark announcement — the Gems freeze is now 9 days out

What happened: Our daily priority check found no new Gemini Spark announcement. Spark remains Task mode inside the Gemini app — the Chat | Task toggle beside the prompt box, multistep tasks running in Google's cloud across connected apps, for Google AI Pro and Ultra subscribers on personal accounts, everywhere Gemini Apps are supported except the EEA, the UK, Switzerland, and Nigeria. The countdown is still the news: personal-account Gems become read-only on November 17 — 9 days from today — and are permanently deleted on December 17, as Google replaces them with Skills. Google auto-converts Gems to Skills, but each skill carries a 100 MB total file limit with oversized attachments dropped largest-first, and several Gem-era capabilities — sharing, Drive files, and use inside Deep Research and Canvas — are still listed as coming later. Because Gemini Skills and Claude skills share the standard SKILL.md file format, skill files you keep yourself are portable between the two assistants.

Who it matters to: Anyone with Gems they actually use, and the window has narrowed from "later this month" to "next week." Nine days is still enough for a calm export; after November 17 the account is read-only, and after December 17 there is nothing left to recover. The portability point matters beyond the deadline: the file format, not either app, is the durable asset — work saved as SKILL.md survives whichever assistant you end up using.

What you can do: Finish the backup this weekend, not next. For each Gem: copy its instructions into a plain SKILL.md file you keep yourself, and download any attachments — especially anything large or in an unsupported format, which will not carry over automatically. Then upload one converted skill into Gemini's Customize → Skills and run it once, so you've proven the format works on your account before the freeze forces the issue.

Learn more: Julian Goldie's guide to Spark's rename to Task, checked against Google's help pages

Also worth knowing

• OpenAI reportedly raised another $11 billion at a $1.25 trillion valuation. Bloomberg reporting relayed by TechCrunch on November 6 puts the unfinished round at $11 billion secured, possibly reaching about $20 billion when it closes in the coming weeks — short of the roughly $30 billion at a $1.4 trillion valuation first sought. All figures come from anonymous sources; OpenAI had not commented when TechCrunch published. Treat as reported, not confirmed.
• Google took over Crusoe's troubled Wyoming data-center project. Crusoe confirmed November 6–7 that it sold the unfinished southeastern Wyoming campus to Google after reportedly raising concerns — Google's, per TechCrunch — about its cost and construction timeline. The $17 billion figure is Google's planned investment to develop the campus, not the undisclosed sale price; the project was originally planned to scale toward 10 gigawatts. Crusoe, valued at $30.9 billion after a $3.9 billion September raise, continues building a separate campus for Google near Amarillo, Texas.
• Ramp data shows business AI spending stalling as Anthropic nears its IPO. Spending-platform data reported by TechCrunch on November 6 shows AI spending growth flattening — landing the same week Anthropic's public listing approaches and OpenAI's reported round came in below its first target. Two data points don't make a downturn, but the "spending only goes up" assumption took a dent this week.
• Hundreds of Artex AI hacking tools were found publicly exposed. Researchers found Artex AI penetration-testing servers reachable on the open web without access controls, per TechCrunch — offensive AI tooling left unsecured, in the same week OpenAI formalized its Hacktron partnership. If your team runs any AI security tooling, check what of it is internet-facing today.
• Ripple-65B Preview is a late catch worth knowing about. Ricursive Intelligence released this 65B-parameter dense model on November 4 under Apache 2.0: pretrained on about 12 trillion tokens, up to 1M-token context, focused on agentic coding, with a vendor-reported 80.52 on TerminalBench 2.1 — self-reported, not independently verified. Weights are on Hugging Face, with hosting on Together AI, GMI Cloud, and OpenRouter; note the card's "up to 1M" context versus roughly 16K on some hosted endpoints.
• ChatGPT Images 2.6 models are now on the independent boards. Artificial Analysis listed GPT-Image-2.6 Sunburst and Flare on its text-to-image and editing leaderboards as of November 7, with Sunburst at the top range for text-to-image — the first independent placement for the models OpenAI shipped November 4, whose per-image costs at low and medium quality run higher than 2.5's at the same size, per OpenAI's own calculator.
• Small tools, small print. Kiro launched Teams credit pooling at $20 per seat per month with 500 pooled credits per seat, and Canva released Affinity for iPad with its core tools free for individuals with a free Canva account — two more examples of AI-era pricing moving to pooled credits and free-core bundles, per this week's change trackers.
• The midterm fallout for data centers continued. TechCrunch's post-election coverage argues the results leave Elon Musk and data-center expansion facing a rougher political road, with voters backing restrictions in most data-center ballot fights — the local-politics constraint on AI build-outs is now a standing story, not a one-off.
• Two small robotics and speech items from the feed. Physicl raised $10 million to make simulated worlds realistic enough to train robots in, and Underdog demoed a speech model that fits in 32MB — the on-device end of AI keeps getting smaller and cheaper to run.

One practical takeaway

Audit your agents' connections this weekend — the seams are where they break. Yesterday's and today's security stories rhyme: Hacktron reached OpenAI's internal systems by chaining a forum flaw into a sign-on into a connected repository, and hundreds of Artex hacking servers were simply left open. Your version of that risk is mundane: the connected apps inside Muse, Gemini, and Alexa, the MCP servers in your coding tools, the integrations you granted in March and forgot. List them, remove the dormant ones, scope the rest to the least access that works — and if you run agents unattended, set them to deny-and-continue rather than prompt-and-hang. Twenty minutes of connection hygiene is worth more than any single model upgrade this week.

Sources

TechCrunch (Amazon's text-based Alexa for Shopping launch — invitation-only U.S. rollout from November 6, search/compare/cart/purchase by text, lists, price tracking, order status and photo search, waitlist for others, wider rollout in coming weeks, the September block of Meta's Muse as reported by TechCrunch, DoorDash's third-party agent order figures and its own text agent, November 6; OpenAI's reported $11 billion raise at a $1.25 trillion valuation per Bloomberg — round unclosed, possibly ~$20 billion total, versus ~$30 billion at $1.4 trillion first sought, anonymous sources, no OpenAI comment, November 6; Crusoe's sale of the Wyoming project to Google — cost and timeline concerns as reported, $17 billion planned Google investment, undisclosed price, 10-gigawatt original scale plan, Crusoe's $30.9 billion valuation and separate Amarillo build for Google, November 6–7; Ramp AI spending data, November 6; Artex exposed servers, November 6; post-midterm data-center and Musk analysis, November 7; Physicl's $10 million raise and Underdog's 32MB speech model demo, November 6 — all via the site's RSS feed and direct reporting); Storyboard18 and AI & Agents Daily Briefing (Muse Spaces availability in all Muse markets from November 7, the Wonder-trained specialized models and Meta's internal 76%-to-90% and 38%-to-47% figures with no published methodology, 10 million Muse users in the U.S. and Canada per Meta's October earnings call, November 8); FourWeekMBA (Amazon's announcement details — invited customers accept terms and link a phone number, requests carry over to the Amazon app, site and Alexa devices, Alexa for Shopping's 400 million+ customers in 12 months per Amazon's Q3 release, November 6–7); Anthropic's Claude Code v2.1.319 changelog via Releasebot, DEV, ToolsNetwork and changelog trackers (background MCP tasks in the terminal and the CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND variable, tasks leaving feature-flag gating for Bedrock/Vertex/Foundry/gateway setups, permissions.denyInsteadOfPrompting, auto-mode changes behind LLM gateways, /claude-api latency-optimize, permission-bypass and sandbox-escape fixes, November 6); Google's ADK Python v2.14.0 release notes via DEV, Emergent and AI Weekly (release dated November 4, tag published November 7; 26 modules and 16 toolset features graduating to stable including execution environments, A2A and evaluation; TeamsRunner, MultiRunConfig, MongoDbMemoryService, generate_agent_card CLI; Google's 23% faster / 9% lower memory claim with no baseline published, November 7); OpenAI's Hacktron announcement via Unite.AI and Silicon Report (partnership focus on third-party service connections, Daybreak Red access and controls, raised bug bounty rewards as reported, November 6) and Hacktron's own September 13 account (the July 25 two-vulnerability chain from the Discourse forum through single sign-on into ChatGPT, Codex and a GitHub integration, ~14-hour confirmed fix, $6,500 bounty paid September 1); Ai2's ToyBlocks release via Unite.AI (November 6 — FoundationWorlds, 85,336 demonstrations, ~1,600 hours, 144 million frames, Apache-2.0 code and ODC-BY data, Molmo2-4B-tb and ManiFlow-tb, 55.5% simulated and 50.0% physical hierarchical success, poor generalization across ManiFlow, π0.5 and MolmoAct 2, physical evaluation limits as Ai2 states them); LLM Releases and FastToolHub (Ripple-65B Preview — 65B dense, 12T tokens, up to 1M context, Apache 2.0 weights on Hugging Face, Together AI / GMI Cloud / OpenRouter hosting, vendor-reported TerminalBench 2.1 score of 80.52, hosted-endpoint context caveats, November 4); Kingy AI's November 7 update (Artificial Analysis listings for GPT-Image-2.6 Sunburst and Flare, OpenAI calculator cost multipliers at low and medium quality); AI Plan Tracker's Week 45 log (Kiro Teams credit pooling at $20/seat/month with 500 credits per seat, Canva Affinity for iPad free core tools, checked November 2026); Gemini Spark guides from Julian Goldie and BetterClaw checked against Google's Gemini Apps Help pages (Task mode requirements, availability exclusions, and the Gems-to-Skills timeline — read-only November 17, deletion December 17, 100 MB per-skill limit). Summaries are original; follow the links for full reporting.