mario03690/dsh-adversarial-review

README 摘要

dsh-adversarial-review Adversarial review for code, MCP configs and system prompts — every finding carries a repro path and confidence; unreproducible ones are marked suspected, never confirmed. What & why Not a scanner: it reviews code/config/prompts YOU provide, it does not attack third-party services. adversarial review probes named surfaces (injection, trust boundaries, secret handling, side effects, idempotency, swallowed errors) — each finding with a repro path and high/medium/low confidence, and anything it cannot reproduce is downgraded to suspected , not presented as confirmed (the same honesty contract as our tables' arithmetic self-check). mcp config audit checks agent/MCP configs for over-broad grants, embedded secrets, unannotated destructive tools. red team prompt attacks your own system prompt and hands back hardening. Pass model to run on a flagship (Claude/Kimi K3/GPT via AllRouter). Start with what can you do — describe your task in any language, get the exact tool and a ready-to-run call. Install Thin config layer only (one @deepseek-ai/dsh-mcp-client row, shipped as cordis.patch.yml ) — no tool code on your machine. Built against the dsh v0.1 developer preview's…

分类