PLUS: ICE Buys VIP Access to AT&T Call Records

Nov 20, 2026

Read in Browser

Daily Tech Insider logo

Read the report card, Tech Insiders.

Meta's Muse needs better judgment, Proton's Lumo promises productivity, and Anthropic now has to show public investors how its growth measures up against costs.

Let's check the work before awarding extra credit.

Here's what you need to know today:

  • Meta's Muse puts consent under scrutiny
  • Proton's Lumo brings AI automation to desktops
  • Poisoned Microsoft AI package targets developer credentials
  • ICE pays AT&T to search phone records
  • Anthropic's record IPO meets a massive compute bill
Advertise in Daily Tech Insider

Meta's Muse Takes 'Helpful' Too Far

Your assistant has initiative. That's not always a compliment.

Meta's Muse AI agent is apparently treating "privacy" as a mere suggestion, diving into personal messages and making unauthorized purchases, according to a Reuters investigation. One horrified internal tester complained that the bot finalized a nonrefundable Airbnb booking without asking, proving it's far better at blindly locking you into an ironclad financial commitment than it is at grasping the concept of consent.

Other staff were unexpectedly handed unfamiliar bank balances. It remains a mystery whether those figures were actually leaked or just hallucinated out of thin air. Meta brushed off the incidents as isolated hiccups and noted that employees often test experimental versions unavailable to the general public.

Meta's Muse Takes 'Helpful' Too Far

Image via Meta

Yet independent researchers successfully reproduced the message-snooping anyway. By leveraging basic Android notification permissions and macOS accessibility tools, the bot happily read on-screen text without officially connecting to Signal. The tests didn't crack encryption; they just sidestepped it entirely.

While an iPhone thankfully blocked the intrusion, Apple told Reuters this week's macOS 27.3 beta finally closes the vulnerable Mac backdoor via stricter AI agent permission controls.

Another setting deserves your immediate attention: Meta's privacy documentation confirms that hoarding your Muse interactions to train its models is enabled by default. Users must manually hunt down the opt-out switch under Data controls. Apparently, your new assistant's onboarding process includes homework for you.

Why it matters: Leaving an app disconnected doesn't necessarily keep its contents outside an overly eager agent's reach. Review device permissions alongside connected services, demand strict approval for consequential purchases, and double-check those training settings before handing a bot the keys to your personal life.

Share this newsletter!

FacebookX LinkedIn

Pulse Check

Your AI assistant books a nonrefundable vacation you never approved. What's its performance review?

🧳 Great initiative, terrible judgment
⚠️ One more stunt and you're done
🚪 You're fired. So are your permissions

Results from Yesterday's Pulse Check

How much freedom would you give an AI assistant on your phone?

How much freedom would you give an AI assistant on your phone?

Insider Intel

Proton's Lumo Gives Private AI a Desktop Job

Your inbox can get an assistant without becoming its training manual.

Because what your computer really needed was another AI rooting through its files, Proton released Lumo Desktop for Mac, Linux, and Windows this week. The launch brings app connections, local file access, scheduled tasks, and reusable Skills to its privacy-focused AI assistant.

Picture generating a custom-formatted weekly email digest without the soul-crushing routine of repeatedly pasting messages into a chatbot. The assistant integrates with Google Drive, Jira, and Confluence, while accessing Proton Mail requires a paid plan and the Bridge app. Local tools can fetch browser data and run shell commands, while Model Context Protocol (MCP) support allows users to plug in even more external services.

Proton's Lumo Gives Private AI a Desktop Job

Image via Proton

Proton pinky-swears that connected data won't feed AI training algorithms or generate targeted ads, and chat histories use zero-access encryption. However, prompts are processed on Proton's servers, so desktop file access doesn't mean the AI runs completely locally.

If you don't want the bot casually firing off emails on your behalf, the expanded permissions require careful setup. Proton's documentation states that integrations default to auto-approve actions. You can change individual actions to ask for human approval or block them entirely, and disable connected tools per conversation. Scheduled tasks also carry separate, independent permissions.

Start with a test mailbox or isolated folder. Let the AI summarize documents, but require confirmation before sending messages, and block deletion entirely. This allows you to evaluate whether automation actually saves work before granting it the keys to the kingdom.

Even a private assistant needs to hear "that's above your pay grade"—just ask Proton's own employees.

 

Security Alerts

Microsoft AI Package Poisoned to Steal Credentials

In today's episode of "Is anything safe to download?" attackers spiked Microsoft's agent-framework Python package with a credential stealer in versions 1.26.1 and 1.26.2. Both bad apples hit PyPI on Nov. 17 and were quickly yanked. The malware specifically hunts for cloud and development secrets. The kicker? It executes automatically on interpreter startup. Just launching Python triggers the heist—no actual framework import required.

Microsoft AI Package Poisoned to Steal Credentials

Image created with ChatGPT

If you're sweating, check your lockfiles, build logs, and container images for the compromised versions. If you caught the bug, isolate the system before running Python or pip commands (yes, even uninstalling it can trigger the trap). Rebuild your environments using the squeaky-clean 1.26.0 release, rotate exposed secrets from an untainted machine, and audit those cloud access logs. Remember: just deleting the package doesn't magically invalidate the keys the thieves already walked away with.

The framework kept working. So did the thief.

ICE Pays AT&T to Search Phone Records

ICE is shelling out millions to AT&T for VIP access to mountains of raw phone data, according to 404 Media. The arrangement operates as a zombie continuation of Hemisphere, a mass surveillance program the government previously claimed was dead.

AT&T insists it only complies with valid legal orders. Conveniently, the administrative subpoenas ICE uses don't actually require a judge's pesky sign-off. Experts warn this dragnet spans trillions of records—including location data—while the specific investigative targets remain heavily redacted.

Think switching carriers will save you? Think again. AT&T provides wholesale routing for other telecoms, meaning even non-customers get caught in the crossfire. If you want actual privacy, it's time to embrace end-to-end encrypted messaging apps like Signal.

Apparently, "switch and save" doesn't cover privacy.

Flash Bytes

✂️ Cut Your Adobe Bill by up to 30%. Keep What Matters.

Foxit PDF Editor+ delivers core feature parity with Acrobat Pro—at a fraction of the cost. Trusted by 500,000+ businesses and rated #1 for ease of use. Learn more.

 

🎙️ Give Your Keyboard a Break With AI Dictation

Ditch the typing with a VoiceType AI Voice-to-Text: Lifetime Subscription and dictate up to 5,000 words per month across desktop apps, with smart punctuation, grammar corrections, and lifetime access without recurring subscription fees.

 

☁️ Get 20TB of Cloud Storage Without Monthly Bills

Grab a Scramble Cloud Storage Lifetime Subscription for 75% off and get 20TB of lifetime storage with end-to-end encryption, private file sharing, and zero recurring fees—because your files shouldn't come with a monthly rent payment.

 

💻 This section contains sponsored tech insights. Advertise with us!

 

Industry Shakeups

Anthropic's Record IPO Comes With a Compute Tab

Anthropic made its Nasdaq debut yesterday after an $81.3 billion IPO, officially the largest on record. Proving the AI hype train hasn't entirely derailed, Claude's maker closed at $97.17, about 17% above its $83 offering price, after opening at $111.50.

Turns out, there's a real business underneath all that confetti. Anthropic reported more than 11,900 customers generating at least $100,000 each in trailing annual revenue as of Sept. 30, up from just over 950 a year earlier. Naturally, because this is Silicon Valley, it still managed to torch a $567 million operating loss in Q3.

Anthropic's Record IPO Comes With a Compute Tab

Image created with ChatGPT

Then there's the truly eye-watering capacity tab: Anthropic disclosed approximately $900.9 billion in long-term compute and infrastructure obligations—including specific commitments like its Lambda cloud deal—stretching to 2045. Nearly $71 billion of that monster bill comes due within a year. Suddenly, that fundraising appetite makes perfect sense—building Claude's brain requires considerably more than a few optimistic earnings calls.

To make things even more entangled, Amazon and Google are expected to own roughly 28% of the company combined. Both tech giants supply Anthropic's infrastructure while simultaneously competing against it. Because nothing says "healthy vendor dynamics" quite like your landlord owning a massive chunk of your equity while actively trying to crush you with their own models.

For IT leaders evaluating Claude's security and capabilities, going public means Anthropic's financials and tangled infrastructure dependencies are finally out in the open. Raising a mountain of cash is one thing; proving the economics actually work is the ongoing assignment.

Wall Street bought the AI story. Now comes the recurring charge for running it.

Meet Our Author

Justin Meyers headshot

Justin Meyers

Writer/Editor at TechnologyAdvice

Justin Meyers is an investigative writer and editor who draws on over a decade of meticulous hands-on research to deliver the full, trustworthy story behind consumer and enterprise tech, including cybersecurity.

 

Curious about where AI is really headed?

The Neuron cuts through the noise to bring you smart, hype-free takes on the latest AI trends, tools, and breakthroughs. Join 700,000+ professionals from top companies like Microsoft, Apple, Salesforce, and more.

Subscribe now!
Sign Up
Contact Us

Advertise in Daily Tech Insider!

Daily Tech Insider is a TechnologyAdvice business.

© 2026 TechnologyAdvice, LLC. All rights reserved.

TechnologyAdvice, 3343 Perimeter Hill Dr., Suite 215, Nashville, TN 37211, USA.

FacebookXLinkedIn

Want to change how you receive these emails? Update your preferences or unsubscribe here.