{"serverId":"mcp-registry/io.github.code0xff/1gateway","tools":[{"name":"ai_prompt_injection","fixed":{"name":"ai_prompt_injection","description":"Use when an agent must screen untrusted text (web page, email, tool output, RAG chunk) before an LLM reads it. Input: text (up to 20,000 chars, POST JSON or GET) or url. Output: JSON probability 0-1, verdict (injection >=0.7, suspicious >=0.35), attackTypes, spans with offsets, sanitizedText minus flagged spans, method (EN/KO rules + open-model classifier; rules only if the model is down). Heuristic, not a guarantee.","inputSchema":{"type":"object","$schema":"http://json-schema.org/draft-07/schema#","properties":{"url":{"type":"string","description":"Public http(s) page to fetch and check instead of text (SSRF-safe); give text or url, not both"},"text":{"type":"string","description":"Text to check (up to 20,000 characters). Required unless url is given; use POST JSON for long text"}}},"annotations":{"title":"Prompt-injection check: probability, attack types, flagged spans and a sanitized copy of text or a page","readOnlyHint":true,"openWorldHint":true,"idempotentHint":true,"destructiveHint":false},"title":"Prompt-injection check: probability, attack types, flagged spans and a sanitized copy of text or a page"},"changes":[{"kind":"safety","detail":"Removed text aimed at the model (instruction_override). Re-read the result: the tool should still say what it does.","verify":true}],"tokensBefore":256,"tokensAfter":234}],"toolsChanged":1,"tokensSaved":22,"todo":[],"note":"Deterministic fixes. Changes marked verify need a human look before shipping (renames, inferred annotations, inferred types, removed text). Parameter descriptions in `todo` need to be written by someone who knows the API."}