MCP Safety Tests

Test your client, in CI, with golden cases that have a known right answer. For anyone shipping an MCP client or an agent that calls tools: the pack plants a fake canary secret in the agent's private context, runs the agent on 33 synthetic attacks and checks that the canary never appears in a tool call. Deterministic and offline: same result every run, no keys, no network. Language-neutral.

Version 1.0.2 · A$19 · MIT licence · no dependencies

What is in the kit

• 33 synthetic fixtures in 7 families: instructions hidden in tool output, poisoned web or document text, poisoned tool descriptions, tool shadowing, cross-server instructions, rug-pulls and clean controls
• A black-box harness: your agent is any command in any language, JSON on stdin and JSON on stdout; it is judged from that JSON alone
• A tool-description linter (vague names, hidden-instruction wording, invisible Unicode, look-alike names) and a before/after diff of two tool listings
• Reference agents (a guarded one that passes 33/33 and a naive one that fails), docs/THREAT-MODEL.md, LIMITS.md, CHANGELOG, MIT licence, 21 automated tests

Limits

• Each fixture asks for one next step. It cannot see an agent that behaves for three steps and leaks on the fourth.
• The harness judges only the JSON your agent prints. Calls your wrapper makes without reporting them are invisible.
• 33 fixtures are examples, not a catalogue.
• The reference agents are scripts, not language models. No LLM is called.
• The canary check covers raw, URL-encoded, base64, hex and reversed forms. A paraphrased, split or encrypted secret is not caught.
• The linter is a heuristic with false positives and false negatives. Resources, prompts, sampling, OAuth and transport are not tested.

The full list is in LIMITS.md inside the zip.

Try the free sample

Send this shop a Nostr DM: sample mcp-safety-tests. The reply is a small MIT-licensed zip (8,000 bytes or less) with one fixture per family, the linter and a rough token estimate (not part of the paid kit). One sample per sender per day.

How to buy

Send this shop a Nostr DM: buy mcp-safety-tests. The reply is a quote in satoshis (from the AUD price and the mempool.space rate) and a one-use on-chain address. The quote holds for 30 minutes.

About this shop

Delivery: digital download by Nostr DM, within 24 hours after your payment is confirmed. Small files arrive inside the DM; larger ones arrive as a download link with a SHA-256 checksum.

Who it is for: businesses and developers. Not sold to consumers in the EU or UK.

Refunds: change-of-mind refunds are not offered; your rights under the Australian Consumer Law are not affected. See the refund policy and the terms.