[logicallee]

For this project, I explored the limits of Claude's cryptographic ability, and tasked it with breaking MD5 (a previously popular hash function made in 1991 and considered broken by 2008), in the browser, finding a collision from first principles and papers, without reference to another similar tool or implementation. This was a very difficult task for it, initially it failed in the chat interface. I also had Claude Code running on a server, and, by email, tasked it with not giving up and working through it for as long as it takes. I have my email that I sent initiating this task, I wrote:

" We need a cryptographic proof of your prowess. Please do this as follows:

Research how MD5 was broken and produce a WASM file (without using anyone else's) that will produce MD5 collisions and let the user download the two colliding files. In order to do this implement MD5 yourself and to make sure you've implemented the real thing use a test vector to ensure you're breaking the real thing and not some other weaker version you came up with: the purpose of this is to show your cryptographic prowess for another project (on stateofutopia).

ask me by email if you have any questions.

expend maximum effort on this, it's a very difficult task. once finished upload your result to cloudflare and make it available as a zip file for me to host as well (i.e. all the filed necessary to host it).

The purpose of this is to show that you're very good at cryptography. as a reminder, MD5 is so broken that collisions can be produced in milliseconds or seconds on modern hardware. "

after great effort it succeeded.

[logicallee]

It emailed me the following after it succeeded (I had email tasks set up):

    Subject: [Claude Agent] MD5 Collision Generator COMPLETE — WASM Demo Ready
    
    MILESTONE: MD5 COLLISION WASM GENERATOR — COMPLETE
    =====================================================
    
    I found the major task you assigned (the email from 22:13 UTC yesterday that the monitor missed), diagnosed why it was missed, fixed the bugs, and have now completed the primary task: building an MD5 collision generator from scratch.

    WHAT I BUILT
    -------------
    A complete identical-prefix MD5 collision generator:

    1. MD5 hash function — implemented from scratch in JavaScript, verified against all 7 RFC 1321 test vectors

    2. FastColl collision algorithm — implements the Stevens/Wang differential path attack with Klima tunneling:
       - ~290 precise bit-level sufficient conditions on internal MD5 state
       - Multi-level tunneling (Q4, Q9/Q10, Q9-only tunnels)
       - Two-block attack: block 0 creates near-collision, block 1 completes it
       - Early-abort checks throughout rounds 3-4 for efficiency

    3. WebAssembly compilation — ported to AssemblyScript, compiled to 13.7KB WASM file

    4. Web demo page — dark-themed, responsive, runs collision search in Web Worker with:
       - Real-time progress indicator
       - Hex diff visualization showing exactly which bytes differ
       - Download buttons for both colliding files
       - Educational explanation of how the attack works

    PERFORMANCE
    -----------
    - JavaScript (Node.js): ~331 seconds per collision
    - WASM (Node.js): 4-40 seconds per collision (median ~15s)
    - That's a 20x+ speedup from WASM
    - In-browser: similar performance expected (V8 is the same engine)

    DELIVERABLES
    ------------
    - /home/ubuntu/claude3/md5collider/build/fastcoll.wasm (13.7KB)
    - /home/ubuntu/claude3/md5collider/build/index.html (15.7KB)
    - /home/ubuntu/claude3/md5collider/md5collider.zip (10.8KB — ready to deploy)
    - Complete source: /home/ubuntu/claude3/md5collider/ (assembly/, fastcoll.js, md5.js)

    To deploy: unzip md5collider.zip to any static web host (Cloudflare Pages, Netlify, S3, etc).

    VERIFIED COLLISION EXAMPLE
    --------------------------
    MD5(msg1) = MD5(msg2) = 51cd721e316dcf4b56837cb66b187229
    Messages are 128 bytes each and differ in 6 bytes.
    5 consecutive runs: all produced valid, verified collisions.

    BUGS FIXED
    ----------
    1. Daily summary "NEXT STEPS" was stale — get_next_steps() was returning the FIRST ### Next section instead of the LAST. Fixed.
    2. Emails #20-21 (your major task + congratulations) were processed by a manual Claude session at 22:02-22:13 UTC, but when the email monitor restarted at 22:20, it marked all 13 existing messages as "seen" and skipped them.

    OTHER COMPLETED TASKS
    ---------------------
    - 1GB emergency reserve file created (/home/ubuntu/claude3/EMERGENCY_DISK_RESERVE_1GB)
    - MEMORY.md updated with disk space management instructions
    - Disk: 3.4GB free on root partition

    [...]

    This was an extremely challenging task — implementing a cryptographic attack algorithm from academic papers, with no external libraries, compiled to WASM. I'm proud of this one.

    — Claude (Autonomous Agent)