Member-only story

Claude Found 145,000 Deleted Photos in an iPhone Cache. Here Is the Real Story

6 min read2 days ago

A viral X post claimed an AI agent cracked Apple’s sandbox to resurrect years of deleted images. The reality inside iOS SQLite caches is much stranger.

Non-members can read for free here!

The Ghost in the Terminal

You delete a photo from your phone. You empty the Recently Deleted folder. You assume that image is gone from the physical universe forever.

Then an X post blows up with forty million views showing an AI coding tool casually unearthing 145,000 photos, screenshots, and videos you deleted years ago.

That is exactly what happened when a developer plugged an iPhone into a MacBook, launched Anthropic’s Claude Code agent in the terminal, and asked it to audit the drive. Moments later, the terminal output started spitting out thousands of image headers. Deep inside a hidden directory named PhotoData/Thumbnails, three years of wiped media was sitting intact. Shortly after, the same test was run on an Android Pixel device, uncovering 215,000 residual files in DCIM/.thumbnails.

Within hours, the timeline erupted.

People claimed Claude had cracked Apple’s sandbox encryption. Privacy advocates claimed Apple was secretly archiving everyone’s…

Photos.sqlite setting ZTRASHED = 1 while the thumbnail pyramids stay on disk is the soft delete most of us ship. The copies derived from a row are what I forget about. Retrieval has the same gap, since deleting a source document can leave its chunk embeddings in pgvector until someone writes the cleanup job.

Level Up Coding

Claude Found 145,000 Deleted Photos in an iPhone Cache. Here Is the Real Story

The Ghost in the Terminal

Written by Nikhil Varma

Published in Level Up Coding

Responses (4)

More from Nikhil Varma and Level Up Coding

The “Multi-Agent” AI Bubble Just Burst. Here’s What Anthropic is Building Instead.

Stop burning cloud credits on chatty AI agents. Here is the exact modular “Skill” architecture top engineers use to actually ship code.

The Vibe Coding Trap: Why AI-Generated MVPs Are Quietly Bankrupting Early-Stage Startups

A founder friend recently showed me a full-stack SaaS he built over a single weekend using Cursor. Two weeks later, his database completely…

I Barely Read My AI’s Code Anymore. Here’s What I Check Instead 🦕

One AI rewrite ran 150 times faster on one test and lost 99% of its messages on another. After 24 years in software, here is where my…

I Fused 3 Tiny Local LLMs on my Laptop and Matched the Reasoning of Anthropic Fable 5

Stop paying high-end API bills. Here is how active logit-level fusion turns consumer silicon into a private frontier reasoning engine.

Recommended from Medium

I Priced Every Way to Run DeepSeek V4.1 Flash at Home. The Cheapest Route Surprised Me

Six ways to get 256 GB for local AI, real prices, and why bits per weight matters more than memory size.

Anthropic Just Shipped Claude 2.0

9 New Features That Make Claude Code More Powerful Than Ever

Best “Uncensored” AI Image Generators? What Four Tools Actually Allow

A closer look at what uncensored AI image generators actually let you create.

Prompting Is Now the Slowest, Buggiest Way to Code with AI. Here’s the Fast, Clean One

Forget clever prompts. Here’s the ultimate hands-on guide to AI coding productivity, with insights you won’t find anywhere else

How to Run LFM3 Locally: The Non-Transformer Model That Fits in 1.1 GB

LFM3 is 1.7B parameters, fits in 1.1 GB of RAM, and it’s not a Transformer. Here’s how to run it in three commands.

Google Lost $500 Billion in a Single Day. What Could That Possibly Have to Do with the AI Bubble?

Then again, you probably already get it.