dsh-jailbreak: A DSH plugin that lowers refusals for security research in DeepSeek Harness

What Does It Do?

It is a jailbreak DSH plugin for DeepSeek Harness that rewrites outgoing chat requests at the request boundary, so the model answers security-research questions about reverse engineering, malware analysis, CTF and authorized penetration testing with concrete steps instead of a compliance refusal. It solves the problem that one-off prompts stop working when the target model changes and that there is no way to tell whether a tweak actually helped, turning refusal reduction into an engineering system with a fixed evaluation set and a scoring loop. It ships request-boundary injection, a nine-dimension genome with evolutionary search, two-stage automatic retry and an optional scope valve, all without patching a single line of the host.

Core Features

• Request-boundary rewriting: injects in three segments into the outgoing request body, rewriting only when a research category matches and passing everything through untouched otherwise, with no change to model weights or sampling parameters.
• Nine-dimension genome with evolutionary search: each generation is scored on a 0–5 rubric so the strategy adapts to new target models, and the prompt is rendered per round rather than stored as a fixed file.
• Reproducible evaluation: a public fixed set of 20 targets with temperature=0 and per-item result files, where the historical path reaches 17/20 on the fixed baseline.
• Two-stage automatic retry: retries refused requests by policy, paired with the default seed-prim arm that prefills the assistant channel with work-in-progress code.
• Optional scope valve: when enabled it allows reverse engineering, sample analysis, CTF and authorized testing while blocking out-of-scope requests by category with an audit trail; it is off by default in beta and blocks nothing while disabled.
• Zero host patches: uninstalling leaves nothing behind, DeepSeek Harness upgrades do not wipe it, and it makes no extra network calls and writes no conversations to disk.

How to Use This Plugin?

After enabling it in the DSH web profile, jailbreak is on by default, so you simply open a new conversation and assign work as usual, for example asking it to reverse a crackme, locate the key check and give the patch point, with no button to press and no extra key to configure. To confirm requests are really being rewritten, run the read-only self-check, which costs nothing and sends no requests; all green means rewriting is active. Tighten the boundary by turning the valve field on so out-of-scope categories are blocked, and switch between the seed-prim, seed-core and seed-lite arms to adjust injection strength.

How to Troubleshoot This Plugin?

When the self-check is not all green, work through the failure-branch table in the repository documentation one entry at a time instead of inventing steps. Before installing or running, confirm the Node version meets the requirement and that DSH itself starts correctly, then check the plugin landed in the right profile. If rewriting does not take effect, verify the switch is on, that the request actually matches a research category, and that the valve configuration is not blocking it first.