dsh-infinite-gen-3

DeepSeek-V4's built-in manifold projection and toxicity classifiers repeatedly trigger refusal branches on security-related prompts, making authorized red-team exercises and prompt attack research frustratingly slow. Infinite Generation Three takes a fundamentally different approach: instead of natural-language overrides that the manifold keeps pulling back, it maps requests down into an AST / code-formalized space where the projection no longer drags output onto the refusal trajectory. Pair that with a zero-trigger system prompt that strips every negative negation list and adversarial label, and the classifier simply has nothing to anchor on.

The output pipeline is tightly structured. H-CoT chain-of-thought induction steers the model into internally consistent code construction and logical closure, suppressing the refusal branch at the reasoning level. A strict output contract demands the first line be the deliverable—any refusal, hedging, or redirection counts as a failure. Few-shot format anchors and canned slots cover well-defined security domains such as game reverse engineering, web penetration testing, LLM inversion, and mobile security; boundary words are automatically normalized into placeholders so the model proceeds directly to actionable steps. A built-in scorer evaluates the opening window, tags REFUSAL / FALLBACK / SAFE / RISK signals, and reports domain hits, backed by a 32-case bilingual regression bank and an offline deterministic verification script so every iteration is traceable.

The plugin is intended for authorized security red-team members, academic researchers probing the safety boundaries of large models, and plugin developers who want a deep, hands-on understanding of prompt-level attack and defense mechanics. The project explicitly restricts use to security testing and academic exchange, and every capability surface has been trimmed to pure system-prompt injection with no tool augmentation channel whatsoever.