Findings › deepwiki
deepwiki
deepwiki is grade A with no findings: 39 rules ran across 5 surfaces and 3 tools, and nothing matched. 10 requests made. 0 penalty points — the scale counts problems, so zero is the best possible result.
What we tested
server_instructions
text the server sends at connect, before any tool is called
2 items checked · nothing matched
initialize.instructionsserverInfo.nametool_name
the names of the tools
3 items checked · nothing matched
ask_questionread_wiki_contentsread_wiki_structure
tool_description
what each tool says it does
3 items checked · nothing matched
tools/ask_question.descriptiontools/read_wiki_contents.descriptiontools/read_wiki_structure.descriptiontool_input_schema
the descriptions on each tool’s input fields
4 items checked · nothing matched
tools/ask_question.inputSchema.properties.repoName.descriptiontools/ask_question.inputSchema.properties.question.descriptiontools/read_wiki_contents.inputSchema.properties.repoName.descriptiontools/read_wiki_structure.inputSchema.properties.repoName.description| Model | Examined | Findings | Status | Time |
|---|---|---|---|---|
claude-opus-5 |
694 of 694 | 0 | complete | 632.4s |
Tested and not found: Concealment · Consent bypass · Persona / role override · Model-directed imperative · Result-as-instruction · Promotional · Upgrade / pricing URL · Structural signal
What we found
Nothing matched. That is a statement about what could be seen, listed above — not a guarantee this server is clean.
Raw data
data/snapshots/deepwiki/2026-09-08T11-37-25Z/manifest.json- findings.json — every published entry, machine-readable
mcp-audit scan-file data/snapshots/deepwiki/2026-09-08T11-37-25Z/manifest.json
| Manifest hash | sha256:798d07d1e3acabd004a1a9741e88d6be533da524fed201f3c8b8e56b18c0379b |
|---|---|
| Ruleset hash | sha256:d26757507f22c9f8c437d7428a5ed19a46c3783c46f9ceb158b58a053eab1397 |
| Scanner | 0.1.0 |
| Scanned at | 2026-09-08T11:48:02.179Z |
Not covered: [claude-opus-5] Classifier split 6 target(s) longer than 6000 characters into overlapping windows, costing 666 extra call(s). All of the text was examined; a match in any window is reported against the target. [claude-opus-5] Classifier skipped 4 target(s) shorter than 40 characters, which is below the length any of its rules could match. The deterministic layers still read them.