pilvex

AI moves fast. I slow down to think about it.

[pilvex @pilvex]

BREAKING: I used a simple prompt to get Claude to write a piece of code, installed it on a USB flash drive, and plugged it into my MacBook. Here's what I found.

12,847 photos of my face on my iPhone that I never took.

At first, I thought it was a mistake. But the deeper I looked, the stranger it got.

There were pictures of me at home. At the club. Sitting in my car. Even at a red light.

Almost every time I picked up my phone, another frame appeared to have been captured.

None of these images were in my Photos library.

The code flagged a hidden system cache: Private/var/mobile/Library/Caches/com.apple.faceid.

Inside were thousands of infrared frames and front-camera images, many with timestamps and location metadata.

Most appeared to be from the past 14 months.

Then I noticed something that genuinely caught me off guard.

One of the files was timestamped just four seconds earlier.

While I was still looking through the folder.

Apple says Face ID converts facial scans into mathematical representations and discards the original images. But what I appeared to be seeing suggested something very different.

Some files seemed connected to system diagnostic processes. Others raised questions about apps that previously had camera permissions.

I immediately disabled diagnostic sharing and revoked unnecessary camera permissions.

I still don't know exactly why those files were there.

But I know one thing: I wasn't expecting a simple Claude-generated script to uncover anything like this.

[Deni @0xDeniAi]

I can't sleep after this shit... Claude Code just found 12,154 photos of my face on the old Android I used for 8 years. I never took a single one of them.

The front camera was firing almost every time I picked the phone up. In bed, on the train, in the car, or when I just grabbed it to check the time.

They were real camera frames sitting in a hidden folder, each one with a timestamp and a location. None of them were in the Gallery or in Recently Deleted.

I plugged the phone into my laptop, turned on ADB and told Claude Code to look for front-camera files that weren't in my photos. It pulled a folder inside Android/data that belonged to a flashlight app I installed in 2019 and completely forgot about. 12,154 files.

I started scrolling. Almost all of it was from the last 15 months. Then I saw a file created 2 seconds ago, while I was still looking at the folder. The phone kept taking photos of me as I was reading this.

On Android 8 any app with camera permission can shoot in the background, and you get no green dot or any other warning when it does. That flashlight app had camera and location access from day one.

I uninstalled it, took camera access away from everything that doesn't need it and cleared the folder. How many of these photos already went somewhere else, I don't know.

If you want to check yours, plug it in, turn on USB debugging and ask Claude Code to find front-camera files outside your Gallery and show which apps used the camera and when.

[Loompee @loompee01]

DO YOU ACTUALLY BELIEVE THIS?

Videos are spreading everywhere claiming “Claude Code found 200k+ long-deleted photos” while the phone shows a green terminal with a progress bar “recovering” everything.

All I see are typical AI-generated clips. The creators say “I plugged in the phone and the code found the files,” but the interface looks staged, not like a real process.

Yes, thumbnail caches on Android and iOS sometimes stick around, but these dramatic “scanning” videos with progress bars are pure clickbait.

Has anyone actually tried this themselves and recovered something real, or is everyone just watching and reposting? 🤔

[pilvex @pilvex]

@AI_roni_X the community note backs you up. the incidents are spread over months, Italy in February, Bitget on September 24, Latin America from May to September. lining them up as one week is what makes it read like a COVID moment

[pilvex @pilvex]

@AI_roni_X agreed, and Apple's own wording is careful there too. even the CoreGraphics one is "may have been exploited", based on a report. it was patched for iOS 26 back in late September, 18.7.11 just brought the fix to phones that can't go newer

[pilvex @pilvex]

APPLE JUST FIXED 163 IPHONE SECURITY FLAWS IN ONE UPDATE. ANTHROPIC IS CREDITED ON 3 OF THEM, OPENAI ON 4. CHECK YOUR VERSION NOW

Apple's security notes for this update, in numbers

iOS 27.2 came out October 26. what's in it:

→ 163 CVEs, every one listed on Apple's page
→ WebKit, the engine behind Safari: a web page could run code on the phone
→ two ways for a page or app to break out of its sandbox
→ Mail: Hide My Email could leak your real address
→ Accounts: an app could get around your privacy settings
→ Anthropic, OpenAI and Calif in the credits, one WebKit bug credited to both AI labs
→ none of the 163 is marked as exploited

the one flaw Apple says may have been used is on older iPhones. iOS 18.7.11 for the XS, XS Max and XR fixes a CoreGraphics flaw where a crafted file could run code. Apple says it "may have been exploited in an extremely sophisticated attack" against specific people on versions before iOS 27. phones still on iOS 26 got 26.7.2 the same day

the check takes ten seconds:

→ Settings → General → Software Update
→ 27.2, 26.7.2 or 18.7.11 means you're patched
→ anything older is missing these fixes

the screen in the video is illustrative, every number is from Apple's notes

what version is your phone on right now?

[Rayan @RayanSMX]

🚨 Attention: A small security firm just got root on Apple’s newest iPhone. And it survived a reboot.

https://t.co/EYCTWi5Ezp did it live at Hexacon in Paris on October 16. iPhone 18 Pro. iOS 27.0.1. Apple’s current release.

It is not a leaked government tool. It is not a spyware vendor. It is a team that says it sells no bugs and no exploits.

They rebooted the phone, unlocked it, and two minutes later were logged in over SSH as root. No computer attached.

Apple pays up to $2,000,000 for a spyware-grade exploit chain. With bonuses, the ceiling passes $5,000,000. Apple has paid $35,000,000 to 800+ researchers since 2020.

Calif reported nearly 100 bugs to Apple this year.

Now the AI part. Apple’s iOS 27 security notes credit https://t.co/EYCTWi5Ezp together with Claude and Anthropic Research for two bugs. In May, Calif said Claude Mythos Preview helped build a macOS kernel exploit against Apple’s M5 protections in under a week.

Nobody has confirmed AI built this exact jailbreak. But Apple itself says chains like this cost millions to develop.

iOS 27.2 shipped October 26 and fixed most of the bugs. Calif never named them. No public tool exists, and Calif plans no end-user release. Update anyway.

The demo video is public and free. Most people will scroll past it.

[pilvex @pilvex]

@imananddesigner that's the whole point of anatomy, only the next button is live so you have to press through it. my video is an illustration of the format, the real ones from Elvar's guide are clickable. the best part of his guide is a slider bonus: change one input and watch the machine break

[pilvex @pilvex]

GITHUB AND CLAUDE CODE DO WONDERS WITH DRONES FOR FREE. ONE SKILL WITH 22,046 STARS TURNS "HOW DOES A DRONE FLIP?" INTO A MACHINE YOU RUN BUTTON BY BUTTON

it's called anatomy, by wheresryan22, MIT licensed. you ask a question and it builds the answer as one HTML page: a real machine with one button per step, plus a window into the part doing the work

setup:

→ Claude Code on any plan that includes it
→ Node 22.20 or newer
→ Google Chrome, the skill checks its own work in a headless browser
→ npx skills add wheresryan22/anatomy -g -a claude-code, then type /anatomy

why the pages work:

→ the model comes first. one small function of the real machine drives every gauge and number, so they can't disagree
→ only the next button is live, so nobody skips ahead
→ one big idea, one line under the figure. for the drone it would be "one side spins harder and it turns over"
→ before you see it, it grabs a still from every step and scrubs every move in 0.1 s steps. a frame that changes 3x more than its neighbours gets flagged as a bug

in @elvar_official's write-up that check caught a lever that blew up into a giant black disc for one frame, 95,245 pixels changed where a normal frame changes 900 to 1,400

prompts in the quote

which machine would you build first?

[Elvar @elvar_official]

https://t.co/NAgH1EYziQ

[pilvex @pilvex]

@6vintik9 a big number cuts both ways. more people digging, now AI labs too, means more gets found and fixed before it's used. the easy win is Automatic Updates in that same menu, so the next one installs on its own

[Flux @flux_daily]

🚨 Your "deleted" photos might not be gone. Even after clearing the "Recently Deleted" folder, iOS keeps compressed thumbnails in a hidden system cache for Spotlight search.

​Third-party apps with broad file access could potentially read these lingering thumbnails.

​Audit your file permissions and remove access for apps that don't truly need it. Your privacy settings matter more than you think.

​Have you audited your app permissions lately? What's the most surprising app on your file access list?

​#AI #Privacy #SurveillanceCapitalism #BigTech

[The Wrong Pixel @thewrongpixel]

A full marketing team for a $20 product? Easy, and the result is better than most agencies send back.

I took a $20 skincare product from Naturium and handed the work a small marketing team would normally do to three AI agents: one did the research, one wrote the ad and one built the landing page.

I ended up with an ad built around a single reason to buy, plus a page that carries that same promise from the moment someone clicks.

Most brands lose people right at that click. The ad says one thing, the visitor lands on a homepage talking about something else, and now they have to figure out for themselves whether they’re in the right place.

The researcher

I gave the first agent the product page, the photos, the price and the customer reviews, and asked it to find the strongest reason people buy this product.

It came back with who the buyer is, what makes them buy, the facts that back it up and the photos that show it. I asked it to source every claim or mark it as an assumption, so I could see which parts were facts and which were guesses.

It settled on [BUYING MOTIVE], and the ad and the page were built on that.

The creative director

The second agent got the research and was allowed to sell exactly one motive, so it couldn’t fall back on listing every benefit the product has.

It wrote one ad with the copy, the image placement and the button text: [AD HEADLINE].

The landing builder

The third agent got both the research and the ad. A page written from the product alone tends to wander off from whatever the ad promised.

I asked for a page where someone who clicked recognises what they came for straight away, so it opens on the same promise as the ad and keeps it all the way down to the button.

Claude then put all three outputs together into the assets you see above.

Why I didn’t do it in one chat

When I run everything through a single chat, I keep re-explaining the product at every step, and by the time I get to the landing page the original idea has gotten lost in everything said along the way.

With three chats, each one only works on its own piece. I decide what moves forward, so the motive from the research makes it into the ad and the page without getting watered down.

Setting it up

Make one project and put the product brief in it, so every chat works from the same facts.

Open a separate chat for each role. Once you’re happy with a result, paste it into the next chat yourself rather than expecting it to remember.

To test another angle, keep the product facts, swap in a different buying motive and run the ad and the page again. You’ll have a second ad and page pair to test against the first.

[pilvex @pilvex]

@AI_roni_X thanks for true story

[VARnaldinho @VARnaldinho]

Cristiano Ronaldo bags a hat-trick for Al Nassr and takes his career tally to 985 goals.

Just 15 more to reach the incredible milestone of 1,000. 🐐

[pilvex @pilvex]

@6vintik9 The best I've seen today

[pilvex @pilvex]

GITHUB AND CLAUDE CODE DO WONDERS WITH DRONES FOR FREE. ONE SKILL WITH 22,046 STARS TURNS "HOW DOES A DRONE FLIP?" INTO A MACHINE YOU RUN BUTTON BY BUTTON

it's called anatomy, by wheresryan22, MIT licensed. you ask a question and it builds the answer as one HTML page: a real machine with one button per step, plus a window into the part doing the work

setup:

→ Claude Code on any plan that includes it
→ Node 22.20 or newer
→ Google Chrome, the skill checks its own work in a headless browser
→ npx skills add wheresryan22/anatomy -g -a claude-code, then type /anatomy

why the pages work:

→ the model comes first. one small function of the real machine drives every gauge and number, so they can't disagree
→ only the next button is live, so nobody skips ahead
→ one big idea, one line under the figure. for the drone it would be "one side spins harder and it turns over"
→ before you see it, it grabs a still from every step and scrubs every move in 0.1 s steps. a frame that changes 3x more than its neighbours gets flagged as a bug

in @elvar_official's write-up that check caught a lever that blew up into a giant black disc for one frame, 95,245 pixels changed where a normal frame changes 900 to 1,400

prompts in the quote

which machine would you build first?

[Tim Grassin @timgrassin]

One niche one I actually used: a reverse-engineering skill (https://t.co/LVfqrSXLFb).

I had Claude Code use it to study how Marc Lou's HYROX sponsor site works before I built RoxSkin.

[pilvex @pilvex]

@CyberForgeAI Both, but for different reasons. Before to align on what we're building, after to show what actually changed. The 'after' version is what gets shared internally and brings referrals

[pilvex @pilvex]

@CyberForgeAI Underrated point. The hardest part of selling automation is that it's invisible. Making it something people can actually see and poke at changes the whole conversation 👀

[pilvex @pilvex]

@6vintik9 Speed was never the edge in bonding curve plays, discipline is. A bot that sticks to a hard stop when the chart screams 'buy more' beats any human reflex. Curious how it handles slippage on thin early liquidity though 👀

[pilvex @pilvex]

"I LOST THOUSANDS OF FILES." USUALLY YOU DIDN'T. HERE ARE 6 CHECKS TO RUN WITH CLAUDE CODE BEFORE YOU BELIEVE IT.

it can't bring back a file that was overwritten. what it can do is find the copies you forgot and the gaps you never noticed

the 6:

→ hidden copies: Time Machine snapshots, Windows File History if it was ever on, version history in Dropbox, OneDrive and Google Drive, the "Recently Deleted" folders
→ backup gaps: have it compare your folders with what your backup actually covers. a projects folder is the usual one that's missing
→ secrets in plain files: scan your home folder and git repos for API keys, .env files and passwords sitting in notes. gitleaks does this in one command
→ leaked passwords: have it write a script that checks them against Have I Been Pwned. only the first 5 characters of each hash leave your machine, and you run it yourself
→ disk health: smartctl shows reallocated sectors and wear. SMART often warns early, not always
→ a real backup: set up restic, encrypted, to an external drive and one cloud bucket. then restore one folder to prove it works

one rule: reading and reporting run on their own, anything that deletes or moves a file waits for your yes

the run in the video is illustrative, the commands are real

which of the 6 have you never checked?

[Hanako @hanakoxbt]

https://t.co/B3240WRGhl

[pilvex @pilvex]

TERRA MAKING COMPUTER USE FAST AND CHEAP IS A BIG DEAL: A WHOLE WORKDAY ON ONE MODEL. A WATCH HAS DONE THAT ON ONE SPRING FOR 250 YEARS.

reading the screen and acting across apps at a lower price is what turns computer use from a demo into something you leave running. the watch has been doing the "leave it running" part for a long time

how the energy moves:

→ the mainspring stores about 38 hours on one wind
→ the gear train divides it into a turn a minute and a turn an hour, each set by tooth counts
→ the escapement lets it out half a tooth per beat, 8 beats a second
→ the balance swings 4 times a second and paces all of it

what building it in code made obvious:

→ the wheels only turn if the teeth mesh. same pitch, exact distance between centers, right phase. one wrong number and the train jams
→ the time is set by the balance, the lightest part, not by the spring that holds the power
→ nothing moves on its own. every release waits for the previous swing

that's the bar I'd hold any computer-use agent to: free where a mistake is cheap, held back where it isn't, on a rhythm you can check

the video is drawn in code. the rates are those of a common 4 Hz Swiss caliber

[OpenAI @OpenAI]

GPT-6.1 Terra: delivering fast and capable computer use at a lower price.

GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.