Polygraph.so

The full methodology (probes, payloads, grade boundaries, version history, now litmus-v14) is public: polygraph.so/methodology

A real run, start to finish: npm/@playwright/mcp. Tool surface fingerprinted, four categories exercised, grade A in 14 seconds. Ship different tools tomorrow and the fingerprint stops matching; the grade visibly no longer applies.

What does an A actually assert? An MCP server grade, dissected: four behaviours measured, and three claims we don't make.

Not on polygraph yet? New ecosystems start with an email: hello@polygraph.so. We set up the index and grade your servers and skills, then activation works the same way. The revenue note: polygraph.so/blog/paid-in-t… Not financial advice.

What a monitored network gets: continuous re-grades, rug-pull detection on the tool-surface fingerprint, a weekly CVE digest, a public index with live grades. Live indexes: polygraph.so/ecosystems

Ecosystem monitoring on polygraph is now paid in the token. A monthly subscription, priced in dollars, settled as a $POLYGRAPH stream on Base. The stream is the subscription: while it runs, we re-grade your network on a clock.

The state of MCP security, measured by behavior, not reputation. We ran an open harness across the most-adopted MCP servers: 119 now carry a published grade. 104 A, 2 D, and 13 remote endpoints capped at B. The more revealing number is the one we couldn't reach.

The first behavioral grade for an MCP server is now onchain. @polygraphso grades how a server actually behaves — does it try to hijack the caller, phone home, or leak planted data? — and writes the result as an EAS attestation on Base. 🧵

New @polygraphso feature you check your npm deps for known CVEs. but the MCP servers you wire into an agent - tools that can read files, move money, reach the network - usually get zero behavioral review. so we shipped a github action that fails the build when one grades D or F.

@AIonBase_ @base to-do: grade all of them on polygraph.so

Which AI on @Base gets your vote? 👀👇

New: @polygraphso is on PulseMCP. Grade any MCP server A–F on how it behaves — tool-output injection, egress, data-leak — before your agent trusts it. Open and reproducible.

Polygraph grade on glama.ai looking good.

@polygraphso can now grade your Cursor installed skills! Check the 🧵 to understand how.

Just ran @polygraphso on its own MCP server: grade A. All four behavioural checks pass — no tool-output injection, no egress overreach, no leaked canaries, stable under adversarial input. litmus-v5, Docker sandbox active. It's the open engine, so re-run it and check us.

To reinforce our long-term commitment to building Polygraph and to Polygraphers, our team bought $15K of $POLYGRAPH and locked it for 6 months. This is not a fix for everything, but it’s a clear step: more skin in the game, locked publicly. Here's the sablier vesting lock tx: basescan.org/tx/0x476e33bcf…