Guides and deep-dives on MCP security — tool poisoning, toxic flows, supply-chain risk and how to scan a Model Context Protocol server before you connect it to your data.
There are two ways to review an MCP server: ask a model to eyeball it, or analyze it deterministically. Seven criteria separate a scanner you can gate a build on from one that only demos well.
Read more →Every client wants a different file, a different top-level key, and a different secret mechanism. Here is the exact path and JSON shape for each, plus why a committed project config installs servers for your whole team.
Read more →Claude Opus 5 is out. More than 2× Opus 4.8 on Frontier-Bench, 3× the next-best on ARC-AGI 3, and near-frontier intelligence at half the price — but not the leader at everything. The real numbers, and the launch-day angle nobody is covering.
Read more →The same MCP server can arrive as an npm package, a PyPI package, a GitHub repo, or a remote URL. Each changes where the code runs, what you can verify, and how it is scanned. A clear side-by-side.
Read more →Until now, checking an MCP server before installing it meant running a CLI or calling an API. The new online scanner removes even that step: paste an npm or PyPI package, a GitHub repo, a live remote endpoint (OAuth sign-in supported) — or your entire client config, and every server in it gets graded at once.
Read more →A beginner-friendly guide to the new free MCP Trust Checker API: what it does, how to get a key, and how to scan any MCP server over a single HTTPS request — with copy-paste examples you can run in two minutes.
Read more →A complete 2026 guide to MCP security: why tool descriptions are a new attack surface, the six main risk categories, and how to actually check a server before you trust it.
Read more →A step-by-step guide to scanning any MCP server for prompt injection, tool poisoning, and supply-chain risk with one free npx command — no signup required.
Read more →How hidden instructions in MCP tool descriptions hijack AI agents — the five attack shapes defenders should recognize, why "looks clean" is not clean, and how to detect poisoning deterministically.
Read more →How the lethal trifecta assembles itself from benign MCP tools, why single-tool scanners miss cross-tool composition, and how a static toxic-flow graph proves the exfiltration primitive before an attacker uses it.
Read more →How to vet npm and PyPI MCP servers before you install them — and how hash-pinning the tool surface and the artifact bytes catches a rug pull, even a same-version republish.
Read more →An MCP server exposes exactly three primitives — tools, resources, and prompts. Here is what each one does, how the client and model use them, how capability negotiation works at connect time, and why tool descriptions are a security concern.
Read more →