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

dsh plugin --profile <your-profile> add github:mario03690/dsh-adversarial-review

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 MCP client config shape (2026-08-13); if a later preview changes it, open an issue for a same-day fix.

Cost, quota, privacy

First heavy call is free (anonymous, no signup); afterwards billed at real upstream cost, reported in every response; failed calls are not charged. Bring an AllRouter key to run any tool on a flagship model at direct rates. The config URL carries ?s=dsh-dsh-adversarial-review — a channel tag identifying the install path, not you.

Disclosure: built and run by the team behind ainetcafe.com. Full bundle: dsh-netcafe. MIT.