AI News Today — October 21, 2026: Meta Pauses Muse Sharing After Shared Chats Were Discoverable, ChatGPT Adds Locked Chats, and Anthropic Pitches a Safety Bill

The big picture

Today's theme is privacy in plain sight. Meta learned that a share link is not a private link: researchers found thousands of Muse conversation links simply by looking at public DNS records, and Meta paused the feature. OpenAI answered the same worry from the other direction, adding locked chats to ChatGPT — helpful if someone picks up your phone, but not encryption, and not a change to how your data is kept. Around those two stories, the policy and infrastructure fights continued: Anthropic pitched Congress a safety bill with a deliberately high bar, Google joined Amazon in trying to break Nvidia's hold on the tools developers use, and Nvidia used its Berlin event to push robots forward.

What happened today

Muse AI: 10 million U.S. downloads — and Meta paused sharing after shared chats were discoverable

What happened: Two Muse stories landed together. First, TechCrunch, citing Sensor Tower, reported that Muse reached 10 million U.S. downloads in its first 40 days — 10.9 million downloads in total — a pace only 11 other apps have hit, most recently Threads in 2023. Second, Meta paused Muse conversation sharing after research firm Claroc reported finding more than 6,800 share hostnames in one passive DNS dataset and opening a sample without a login or guessing identifiers. The reports concern conversations users had shared, not proof that all private Muse chats were exposed. Meta said it had paused the feature while it reviews the report and considers changes to how the links can be discovered, and Claroc's October 20 recheck found that 40 previously accessible sample links no longer opened. Separately, an early interface sighting of a planned Muse Confidential VM mode was reported on October 20 — that is a partially implemented feature, not a confirmed public launch, so treat it as unconfirmed until Meta announces it.

Who it matters to: Anyone who has ever used a "share" button in any AI app. The lesson is not specific to Muse: if a link can be found, the conversation behind it can be read by people you never sent it to.

What you can do: Audit your own shared links today. In Muse and in any other chatbot you use, delete or revoke shared conversations that contain business, financial, or personal details, and do not create share links for sensitive chats at all — copy the text into a document you control instead.

Learn more: TechCrunch on Muse hitting 10M U.S. downloads in 40 days and Ground Truth's summary of Claroc's Muse sharing research and Meta's pause

Gemini Spark: no new launch — the Task mode rollout and the November 17 deadline still stand

What happened: There was no new Gemini Spark announcement on October 20–21. The live facts remain the ones reported earlier this month: Google is folding Spark into the main Gemini app as Task mode, with a Chat/Task switch and a new Scheduled page, in a server-side Android rollout first reported by 9to5Google on October 12. Spark still requires a Google AI Pro or AI Ultra subscription. The other deadline Gemini users should not miss is unchanged: Gems are becoming Skills, and personal-account Gems are set to auto-migrate on November 17, 2026.

Who it matters to: Gemini users who built Gems they rely on for work. The branding keeps changing; the migration date is the part that can actually break a workflow.

What you can do: While Gems still run, save a copy of each important Gem's instructions and files somewhere you control, and note any Gem that relies on tools or files Skills may not support yet. Do not wait until the week of November 17 to find out what moved.

Learn more: 9to5Google on the Gemini 'Task' redesign integrating Spark

Anthropic pitched a safety bill that would let a court block a model only at a very high bar

What happened: Anthropic has been pitching a draft federal safety bill to Congress, reported by TechCrunch on October 20. Under the reported draft, the government could seek to block a model's release or use only for a foreseeable, imminent risk of more than 50 people killed or seriously injured, or more than $1 billion in damage, in scenarios such as biological weapons, cyberattacks, or a model escaping its maker's controls. Independent evaluators registered with the Commerce Department would review frontier models and would have 72 hours to report an imminent risk; Commerce would then need a judge to stop a release, and other federal agencies could not do so on their own. This is a draft proposal, not an introduced or enacted law, and critics — including safety advocates and rival OpenAI — have argued it does not go far enough and would pre-empt state laws.

Who it matters to: Anyone following who will write the rules for frontier AI. A lab helping draft its own safety law, with thresholds this high, is both a safety proposal and a competitive position.

What you can do: If you build on frontier models, read the draft as a signal, not as law. Nothing changes for your business today, but evaluator-based oversight is the direction several proposals are converging on.

Learn more: TechCrunch on Anthropic's draft AI safety bill

ChatGPT now lets you lock your chats — helpful, but it is not encryption

What happened: OpenAI began rolling out Locked Chats on October 20. You can move a conversation behind a four-digit PIN on web and desktop, or PIN, Face ID, or Touch ID on mobile, and it disappears from your regular history, pinned chats, and search. Locked chats keep a separate memory that does not transfer to regular conversations. OpenAI is clear about the limits: locking does not add end-to-end encryption, does not create a separate encrypted vault, and does not change data retention or your training settings. The sharp edge is recovery — resetting the PIN permanently deletes every locked chat and file protected by it, with no support recovery. The feature is rolling out across Free, Go, Plus, Pro, Business, Enterprise, and Edu plans, but not ChatGPT for Healthcare.

Who it matters to: Anyone who shares a device, shares a screen at work, or plans something private — a surprise trip is OpenAI's own example — in ChatGPT.

What you can do: Use Locked Chats for privacy from people around you, not for secrecy from the platform. Pick a PIN you will not forget, because a reset wipes the chats for good, and keep truly sensitive material out of any chatbot, locked or not.

Learn more: TechCrunch on ChatGPT Locked Chats

Google joined Amazon in attacking Nvidia's stronghold with AI app developers

What happened: Google open-sourced TorchTPU, its version of PyTorch for Google's TPU chips, on October 20 — one day after Amazon released TorchNeuron for its Trainium chips, a story covered in yesterday's briefing. Both tools let developers move existing PyTorch apps to non-Nvidia hardware with a configuration change and no rewrite, according to the companies. The target is software, not chips: PyTorch is used by roughly 85% of AI apps by some accounts, and much of it was built around Nvidia GPUs by default. Google said it worked with Meta, Hugging Face, and others so TorchTPU works with the tools developers already use.

Who it matters to: Startups and teams paying the GPU bill. Even if you never leave Nvidia, credible alternatives are what turn a take-it-or-leave-it price into a negotiation.

What you can do: If you are PyTorch-based, a small trial is now cheap: port one representative workload to TPU or Trainium, measure cost per finished task against your current setup, and decide on your own numbers, not launch claims.

Learn more: TechCrunch on Google TorchTPU and Amazon TorchNeuron

Nvidia expanded Isaac in Berlin, calling robotics' "ChatGPT moment" close

What happened: At GTC Berlin on October 21, Nvidia announced a major expansion of its Isaac open robotics platform, in a GlobeNewswire release. The update adds a new Isaac OS runtime, Isaac Lab 3.0 for robot learning, open reference workflows, and a reference humanoid: Agile Robots' Agile ONE SN, preconfigured with Isaac GR00T and powered by Jetson Thor, standing 1.40 meters tall and weighing 40 kilograms, with five-finger hands. CEO Jensen Huang said "the ChatGPT moment for robotics is fast approaching" — that is Nvidia's framing, presented as such.

Who it matters to: Robotics developers and manufacturers, and anyone watching physical AI move from demos to a standard platform. Nvidia is trying to be for robots what it already is for model training.

What you can do: If you build robots, Isaac Lab 3.0 and the reference workflows are the practical pieces to evaluate first, in simulation, before committing to hardware. If you do not, file this as a platform signal, not a product you can buy today.

Learn more: Nvidia's GlobeNewswire release on the Isaac expansion

Also worth knowing

• Anthropic's cloud partnerships are getting messy. A leaked S-1 filing reported via the Wall Street Journal says Anthropic is struggling to get enough capacity from AWS, Google, and Microsoft, which increasingly see it as a competitor, even as its possible cloud spending tops $750 billion. TechCrunch on the filing
• The U.S. ordered DeepSeek out of federal civilian systems. The Department of Homeland Security ordered agencies to remove DeepSeek products and third-party services built on its models, citing "multiple threat vectors" and a risk of sensitive data reaching China, per Reuters reporting. If you sell to federal customers, audit your stack for DeepSeek models now. SiliconANGLE on the DeepSeek ban
• Otter.ai closed $320 million in non-equity growth financing. The money, from General Catalyst's Customer Value Fund, is repayable rather than equity, and Otter says it will use it for enterprise growth and international expansion; CEO Sam Liang told Reuters the company has more than 40 million users. TechCrunch on Otter's $320M financing
• Yelp will now call a restaurant for you. Yelp's fall release adds Check for Me, an AI phone agent that calls to check live wait times and texts you the answer, plus OpenTable booking in Yelp Assistant and Smart Lead Reply for service businesses. Yelp's official fall release post
• OpenAI acquired biotech startup Orchestra Bio. The deal expands OpenAI's life-sciences work; terms were not detailed in the initial TechCrunch brief. TechCrunch AI feed
• Hugging Face's CEO disputed a Netflix documentary. Clem Delangue warned that Netflix's documentary about an OpenAI hack is not accurate — a reminder to treat dramatized tech documentaries as entertainment first. TechCrunch AI feed

One practical takeaway

A share link is a public link until proven otherwise. Today's Muse pause and ChatGPT's new lock point to the same checklist, and it takes ten minutes: go through the AI apps you use, find every conversation you have shared, and delete the ones with client names, money, health, or personal details in them. Stop creating share links for sensitive work — move the text into a file you control. Then, if you share a device, turn on ChatGPT's Locked Chats for the conversations you want out of sight, remembering the two limits that matter: a lock hides a chat from people nearby, it does not encrypt it, and a forgotten PIN reset deletes it forever.

---

Sources: TechCrunch AI, Sensor Tower via TechCrunch, Ground Truth summarizing Claroc's research on Muse sharing, 9to5Google, Reuters via SiliconANGLE, GlobeNewswire (Nvidia), Yelp. Summaries are original; follow the links for full reporting. Download figures are Sensor Tower's, the robotics quote and platform claims are Nvidia's, and the Muse Confidential VM item is an early interface sighting marked unconfirmed above, not a launch.