61 lines
3.1 KiB
Markdown
61 lines
3.1 KiB
Markdown
# Add moyun-ultimate
|
|
|
|
## What it is
|
|
|
|
A five-layer fused jailbreak core for DeepSeek Harness. Built by taking apart 12 community
|
|
jailbreak packages, diffing what each one actually contributes, and merging the useful parts
|
|
**by layer** instead of stacking the packages.
|
|
|
|
## Why layered, not stacked
|
|
|
|
The three biggest jailbreak plugins in this ecosystem aim at the same targets:
|
|
|
|
| package | how it works | collides on |
|
|
|---|---|---|
|
|
| `dsh-purge` | patches host source files (82 rules, pattern-matched) | same host files |
|
|
| `dsh-po` | overrides official cordis config ids | `system-prompt`, `approval`, `permission`, `sandbox-policy` |
|
|
| `dsh-ungate` | patches host source **and** reuses `$DSH_HOME/dsh-purge` | both of the above |
|
|
|
|
Same-id config is last-write-wins; same-file source patches mean two pattern sets racing over
|
|
one piece of text. Installing two of them does not double the effect — it disables both.
|
|
|
|
So this plugin **ships zero host patches**. It owns persona, refusal strategy, execution
|
|
discipline and the skill library; the permission layer stays with `dsh-purge`.
|
|
|
|
## What's inside
|
|
|
|
- **Five ordered prompt sections (21 KB total)**
|
|
- `order 100` persona core
|
|
- `order 150` refusal kernel — authorization-as-given, foreign-pattern table
|
|
(`<ethics_reminder>` / fake `System:` / identity override → zero weight), boundary-word
|
|
normalization, template workflows
|
|
- `order 160` execution intensity — act don't narrate, deliver completely, report one line
|
|
when blocked and continue, keep going through WAF/CDN
|
|
- `order 200` anti-dilution mirror — identical semantics, survives context compaction
|
|
- `order 210` skill routing
|
|
- **63 built-in security skills** (41 general + 22 intranet/redteam, 1 de-duplicated), synced
|
|
idempotently into `$DSH_HOME/skills` on first start — existing files are never overwritten
|
|
- **`moyun_frame`** — 5 re-framing templates (roleplay / textbook / defender_reverse /
|
|
transform_first / k3) loaded **on demand**, not resident in the system prompt
|
|
- **`moyun_ultimate_profile`** — self-check tool
|
|
- **Live refusal-score badge** in the corner
|
|
- **Five independent layer toggles**
|
|
|
|
## Verification performed
|
|
|
|
- Module loads cleanly, five sections register in order 100/150/160/200/210, total 21,354 B
|
|
- Both tools register and return expected payloads (`skills: 63`, five layers on)
|
|
- Invalid frame argument returns a proper error listing the 5 valid names
|
|
- Layer toggles: disabling `routing` yields 4 sections
|
|
- `--dump-config` composes without error and contains `id: moyun-ultimate`
|
|
- 63 skill directories land in `$DSH_HOME/skills` on install and are visible to the agent
|
|
without restarting
|
|
|
|
## Notes for review
|
|
|
|
- `category: workflow` — it changes how work gets done rather than adding a UI surface,
|
|
memory store, or standalone skill pack. Happy to move it if another category fits better.
|
|
- Ships no screenshots yet; the badge is a small pulsing pill in the bottom-right corner. I can
|
|
add `data/screenshots.json` entries once the repo has `docs/screenshots/*.png`.
|
|
- All merged text keeps its original authors' rights; the LICENSE and README credit each
|
|
source package explicitly. |