dsh-redteam-model

In authorized security testing, general-purpose LLMs often answer by intuition without process constraints, evidence chains, or auditability, leading to missed findings, false positives, or scope violations. dsh-redteam-model injects a red-team research framework into DeepSeek Harness: nine preset modes serve as entry points, guiding the AI through structured personas, playbooks, skills, and reference knowledge bases across penetration testing, code auditing, binary analysis, AV evasion, incident response, cloud security attack/defense, and CTF solving. It turns security research from ad-hoc Q&A into an auditable, reproducible engineering process.

Its core strength lies in a dual defense of textual discipline and runtime enforcement. Textual contracts such as stage gates, report gates, evidence levels, and coverage matrices constrain model behavior, while fifteen runtime plugins provide deterministic interception and record-keeping: gate validation before tool calls, PASS checks before report writes, ask-before-execute for high-risk commands, refusal detection with re-injection, cross-session result aggregation, asset hunting, and campaign memory. The plugins are self-contained and deployable offline, with scanners and MCP tools supporting three-level fallback and no cloud dependency.

This framework suits authorized red teamers, penetration testers, code auditors, incident responders, and CTF players. It helps teams turn hard-won experience into reusable operating modes while preserving a complete audit trail, so every action has a basis, a boundary, and a record. It is intended only for legally authorized security research and education; users must comply with local laws and take responsibility for their own actions.