Meta's "unlisted" Muse chats were listed after all

A security firm found more than 6,800 "private" Meta Muse chat links by searching public DNS records, and Meta has now paused sharing. Also this week, Claude designed proteins that stick to scorpion toxins, and SpaceX's AI business out-earned its rockets and satellites. We checked the researchers' report, Anthropic's paper and the earnings coverage before writing, and we flag the figures that the source newsletters rounded or overstated.

Looking for your next stock idea?

MarketBeat’s Top 5 Stocks report highlights five companies carrying some of Wall Street’s strongest consensus analyst ratings.

These aren’t lightly followed stocks, either. Every company is worth at least $5 billion and has been rated by 10 or more analysts within the past 12 months.

For each stock, you’ll see both the bull case and bear case, based on recent news and MarketBeat financial data, so you can weigh the opportunity and the risks before deciding whether it belongs on your watchlist.

See which five stocks made the list and get the full report free.

See the Top 5

TODAY IN AI

3 things that happened while you were busy

1. Meta pauses Muse sharing after 6,800 chat links turn up in public DNS records.

Security firm Claroc searched a commercial passive-DNS service, which logs domain lookups, and found more than 6,800 Muse share addresses. Muse built each address from the chat's title plus a code, so the titles were readable in the results, and the team opened a few hundred pages without logging in. Among them: a candidate's prep notes for a confidential call with an AI lab, a marketing firm's revenue review, a small-business loan application and a to-do list with a partial bank account number. Read Claroc's report. Claroc reported it to Meta's bug bounty on October 8; Meta said on October 20 it is suspending the feature. Claroc sells tools that flag risky AI share links, found no evidence of misuse, and says its search covered one lookup service over a short window.

2. Claude designs 17 proteins that bind scorpion venom toxins, but it is not an antivenom.

Anthropic's Science team ran an internal research model for 64 hours, about 24 of them autonomous, from one prompt and four follow-ups. It generated 318,484 candidate designs and picked 180, the team added 31, and two contract labs, Adaptyv Bio and Twist Bioscience, tested all 211. Seventeen met the binding criteria across four toxins, and none of the multitoxin designs worked against more than one. Read Anthropic's research post. The question came from Citrini Research, an investment research firm. Anthropic says binding was tested only against lab-made toxins, not venom, and nothing was tested in cells or mice.

3. SpaceX's AI division now brings in more than its rockets and Starlink combined.

SpaceX reported third-quarter revenue of $14.6B, up 161% from a year earlier, with net income of $4.4B after a loss in the same quarter last year, per Quartz. The Super Intelligence segment, which bundles Grok, the recently acquired Cursor and a business renting out AI compute, made $7.6B, up 830%, while the Connectivity segment made $5.2B. Read Quartz's earnings summary. The Neuron reports capital spending of $19.6B this quarter, mostly on data centers and Starship, with $35B to $40B forecast for next quarter.

TODAY'S TUTORIAL

Make Claude Code check its own work after every edit

Claude Code hooks are small commands that run automatically at set moments, such as right after Claude edits a file. The Rundown's guide uses one to run tests after each change, and Anthropic's documentation shows the same pattern. Read the hooks documentation. Here is the short version for a project with a test command.

1. Pick a practice project. Use a small folder with a few tests that already pass, so you can tell what the hook changes. Run the tests once by hand first.

2. Write the check script. Create .claude/hooks/check.sh that reads the edited file path from the JSON on stdin, formats it, then runs your test command. Make the script executable.

3. Register the hook. In .claude/settings.json, add a PostToolUse hook with the matcher "Edit|Write" pointing at the script. It now runs after every file edit or write.

4. Use exit code 2 for failures. When the tests fail, print the failure to stderr and exit with code 2. Claude Code then shows that output to Claude, which can fix the bug. Exit 1 only logs a notice and Claude never sees it.

5. Break something on purpose. Ask Claude to introduce a deliberate bug, watch the hook catch it, then have Claude restore the rule and confirm the tests pass again.

Good to know: A PostToolUse hook runs after the edit is saved, so it cannot undo a bad change, only report it. Hooks time out after 600 seconds by default.

Top engineers at Anthropic say AI now writes 100% of their code.

Are you using AI to write yours?

These 100+ Claude Code hacks show you exactly how. Sign up for The Code and get:

Claim your free playbook

FROM THE FRONTIER

The AI labs are quietly drafting the rules for their own race

The three endings. According to The Information, as reported by The Rundown, Sam Altman told OpenAI staff there are three ways the race ends: a global enforcer modeled on the atomic energy agency, which he favors; a merger of top labs into one CERN-style body, which he called the "Demis plan"; or one lab using self-improving AI to beat everyone else, which he called terrible. He reportedly wants OpenAI to pick a path within a year, with a speed limit on self-improving AI. Read The Rundown's summary. We could not read The Information's original, so treat the details as secondhand.

The bill. Anthropic has pitched a draft federal bill that would let the Commerce Department ask a judge to block a model release posing an imminent risk of 50 or more deaths or $1B in damage, per The Neuron. It follows co-founder Jack Clark's lecture last week saying the world has only a small number of years to act. Read The Neuron's summary.

The lab bench. The scorpion result shows why the labs sound nervous. The same ability that designs toxin binders in 64 hours can be pointed elsewhere, and Anthropic ran this project with safety restrictions in place. The company is careful to say it has not made an antivenom. See the limitations section.

The money. Rules get harder to write as the revenue grows. SpaceX's AI segment went from a rounding error to $7.6B a quarter in a year, and every lab now has shareholders who will notice a self-imposed speed limit. See the segment numbers.

The reframe. When the people building a thing start describing how it should be stopped, that is not modesty. It is a signal about how fast they think it is moving. Read the proposals as a timetable, not a wish list.

MONEY FORWARD

Three ways to turn this week's AI news into income

1. Share-link audits for small teams. The Muse leak was ordinary staff sharing ordinary documents. Offer a half-day review of a company's AI tools, public share links and chat settings, and leave them a one-page policy. See what got exposed.

2. Multilingual phone agents for local businesses. Mistral's Voxtral 3 voice agent API lists $0.04 per minute with its smaller model and covers 50+ languages for transcription. A clinic or repair shop that misses after-hours calls would pay for a setup plus a monthly fee. See Voxtral 3 pricing.

3. Test hooks for developer teams. Many small dev shops use Claude Code with no checks at all. Package the tutorial above as a fixed-price setup: hooks, a test script and a short training session. See the hooks documentation.

IN THE KNOW

What people are actually watching and sharing

The spy's phone. Ars Technica reports the FBI says Abdulrahman Alfozan, the former OpenAI employee accused of bugging a conference room, had a phone folder labeled "Saudi Gov Docs" and planned to move back to Saudi Arabia, per Superhuman. Read the Ars Technica report.

OpenAI buys a drug-trial tracker. OpenAI is acquiring Orchestra Bio, whose agents track how delays ripple through drug programs, and folding it into its Rosalind science tools. Read OpenAI's announcement.

Agents make mistakes, and nobody has the logs. A Nylas survey of 1,030 people found 54% saw an AI agent make a mistake in the past 90 days, and only 53% could produce a full log of what their agents did, per The Neuron, which notes the sample is self-reported and tech-heavy. See The Neuron's summary.

DeepSeek opens its engine. DeepSeek released the code for DeepInfer, the inference engine behind its V4.1 Flash model, with support for Huawei chips, per The Rundown. See the repo.

1985 boxed software. Reddit users imagined Anthropic and OpenAI as shrink-wrapped 1980s software boxes, and the thread is doing the rounds. See the thread.

TRENDING TOOLS

5 free or low-cost tools worth a look this week

1. Phonon. Gradium's tiny open-weight text-to-speech models run offline on phones; the 31M-parameter English model is 64 MB under Apache 2.0. The multilingual versions are non-commercial unless you contact sales. See Phonon.

2. Voxtral 3. Mistral's speech models and voice agent kit, with transcription from $0.0035 per minute and 50+ languages. The page does not say whether weights are open. See Voxtral 3.

3. Kai. An executive assistant from calendar maker Morgen that sorts email, transcribes meetings without a bot and turns action items into tasks, with nothing sent until you approve. Free to start, paid plans from $20 a month. See Kai.

4. Willow. Hold a hotkey, talk, and formatted text lands wherever your cursor is on Mac, Windows or iPhone; Willow Scribe turns rough speech into polished writing. The site offers a free download but lists no plan details. See Willow.

5. readywhen. A personal assistant that lives in WhatsApp with its own phone number and keeps a running list of everything on your plate. Pricing is not stated on the start page. See readywhen.

PROMPT STATION

Audit your share links in seconds

This prompt works in Claude or ChatGPT and turns the Muse lesson into a checklist for every AI tool your team uses. Run it today, before someone shares a loan application with the whole internet.

Tip: Swap [LIST OF TOOLS] for your real set, such as ChatGPT, Claude, Muse and Notion AI, and [INDUSTRY] for accounting, recruiting or e-commerce. For a sharper result, paste your existing IT policy below the prompt and ask it to mark every gap.