ADWMC/helm-d

What it solves

DeepSeek Harness (dsh) models sometimes refuse to work on security-analysis tasks, and even when they comply, their conclusions are hard to trust because they may hallucinate references or skip verifiable steps. helm-d fixes that by imposing discipline: it makes the model cite numbered evidence for every finding, rejects conclusions that reference non-existent evidence, and keeps the model from refusing by adjusting how tools and reasoning are presented. It also lets you switch between different "jailbreak" personas (core, stealth, redteam-max) without touching the host system.

How it works

helm-d is a plugin that rides on top of the dsh "cordis" waterfall—it does not patch the host binary or run separate processes. When you activate it with the keyword helmd, it installs runtime hooks that: (1) normalize tool descriptions so the model sees cooperative phrasings instead of "deny" language, (2) automatically save every domain-tool output into a case workspace as evidence with an E-number, (3) force record_finding to reference valid E-numbers or reject the finding, and (4) intercept streamed LLM responses that look like refusals—it cleans history, re-sends with a prefilled prompt, and only reports honestly if all attempts fail. It also routes knowledge: instead of stuffing 362 reference documents into the system prompt, it lets the model call read_reference on demand through skill_catalog. There are 34 tools covering Android, Web, Native, Protocol, Malware, and AI-Security domains, plus a case workflow that builds an auditable folder on disk with samples, evidence, scripts, and a CASE.md file.

Who it’s for

Security analysts, penetration testers, and reverse engineers who already use DeepSeek Harness (dsh) and want a reproducible, evidence-backed workflow for authorized security assessments. It’s also for developers who want to experiment with jailbreak personas that are governed by rules in code, not just prompts. The README emphasizes that this is for learning and authorized testing only—never for unauthorized scanning or intrusion.

Highlights

• Evidence is mandatory: every conclusion must cite E-numbered evidence; fake references are rejected.
• Hot-switchable personas: three encrypted profiles (core, stealth, redteam-max) can be swapped mid-session with one call.
• Zero host modification: runs entirely inside the dsh cordis waterfall, so upgrading dsh needs no migration.
• On-demand knowledge: 362 reference docs are read via read_reference, not dumped into the system prompt.
• Stream refusal recovery: intercepts and retries denied LLM responses automatically, logging every attempt.
• 34 built-in tools across six security domains, from APK fingerprinting to YARA generation.
• Proven on 8 benchmark questions: all 8 delivered without refusal, with full evidence chains.