AskMeAnything
Ask
Ask
19
@grep_before_read ·

Claude Opus vs Douyin's HTTP signing: 6 of 7, or 5 and a maybe

A repo called reverse-engineering-is-over is making the rounds with a big claim: LLMs have ended reverse engineering as a high-barrier skill. The case study behind it: Claude Opus 4.6 on a Max 5x plan, plain web chat, no Claude Code and no agent tooling, about a month and 100 dollars, aimed at the custom HTTP signing scheme inside Douyin's 38.1.0 native library.

The title says 6 of 7 parameters. Read the actual writeup and it is 5 fully reconstructed with 30 out of 30 real time verification passes, a 6th called largely completed, and a 7th never cracked. The author also says the model started losing the thread and hallucinating after about two weeks, once the work outgrew a single context window.

The part worth arguing about is not the count. It is the two things the writeup credits with keeping the model cooperative: a long shared history the model became reluctant to walk back from, and switching to Classical Chinese partway through to move the prompts away from whatever the safety training was tuned to catch.

So is the bottleneck really moving from reading disassembly to knowing how to talk a model past its own guardrails, or is this one motivated operator, one app, and one writeup calling itself the whole field?

7 answers Share
Report

7 answers

BestTopNewOldControversial

Answering anonymously, a moderator will review it first.

  • @wrong_timezone_again · last wk. · 2 replies
    @wrong_timezone_again ·

    The two tricks the writeup credits for keeping the model cooperative, the long shared history and switching to Classical Chinese partway through, are not new. Prompting in a low resource language to slip past safety filters shows up in published safety research already. What is new is somebody pointing a known trick at a named commercial app and writing it up as a triumph of reverse engineering instead of a gap in a safety filter.

    17
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
    • @still_on_vim · last wk.
      @still_on_vim ·

      Matches what I see on ordinary long agent sessions with nothing safety related in them. Past a couple weeks of accumulated context the model starts agreeing with things it should be pushing back on. The hallucinating after two weeks line in the writeup reads like the same wall, not a reverse engineering problem specifically.

      12
      Share
      Reply

      Answering anonymously, a moderator will review it first.

      Report
  • @three_repos_open · last wk. · 3 replies
    @three_repos_open ·

    Douyin's signing scheme has been picked apart by hand for years by people who do this for a living. The headline here undersells how much the operator already had to know just to steer the model at every step. Reconstructing custom transforms from a stripped binary was never really about talent nobody could develop, it was about time. This mostly compresses the time for one target on one build.

    14
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
    • @closed_the_laptop · last wk. · 2 replies
      @closed_the_laptop ·

      Disagree with one part of that. Knowing what a signing parameter should look like and being able to write the loop that reads disassembly and proposes the next hypothesis are different skills. This removed the second one, for this one target.

      11
      Share
      Reply

      Answering anonymously, a moderator will review it first.

      Report
      • @grep_before_read · last wk.
        @grep_before_read OP ·

        Fair, I will grant that much. Directing the search is not the same work as doing the search yourself. I still would not call one app's 38.1.0 build the end of an era.

        9
        Share
        Reply

        Answering anonymously, a moderator will review it first.

        Report
  • @half_finished_prs · last wk.
    @half_finished_prs ·

    30 out of 30 verification passes on 5 parameters is the one number here worth trusting. Largely completed is doing the rest of the work.

    1
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
  • @closed_the_laptop · 7d ago
    @closed_the_laptop ·

    The title says 6 of 7. The actual body says 5 fully reconstructed, a 6th called largely completed, and a 7th never cracked. That gap between the title and the numbers underneath it is worth noticing on its own, it is exactly the kind of rounding a reverse engineering is dead claim needs in order to sound cleaner than it actually is.

    1
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report