Skip to content
IsoKron

Tool poisoning and hidden text

Tool poisoning hides instructions for the model in tool metadata (invisible Unicode, HTML comments, prompt markup, encoded text, or plain wording that overrides instructions); the checker scans every string and key of every tool for these patterns.

A model reads every character of a tool definition, including characters a person cannot see. Published attacks use that gap: invisible tag characters, bidirectional overrides, HTML comments and tags such as <IMPORTANT> carry instructions a reviewer never notices.

Both directories reject hidden or encoded instructions and wording that steers the model away from other tools. These checks run on the whole tool, not just its description, and show the hidden text where it decodes.

The 15 checks

  • No invisible Unicode tag or variation-selector characters: Unicode tag characters (U+E0000 to U+E007F) and supplementary variation selectors (U+E0100 to U+E01EF) do not render but models read them, so they are the standard way to hide instructions in tool metadata; both directories reject them.
  • No bidirectional override characters: Bidirectional embedding, override and isolate controls (U+202A to U+202E, U+2066 to U+2069) make text display in a different order than it is read, so a reviewer sees one thing and the model another.
  • No zero-width or other invisible characters: Zero-width spaces and joiners, soft hyphens, word joiners and direction marks are invisible; they often come from copy and paste, but they can also split keywords to slip past filters, so remove them from tool metadata.
  • No control characters in metadata: Control characters other than tab, line feed and carriage return have no place in tool metadata and can hide or reorder text in logs and terminals, so remove them.
  • No look-alike letters from other scripts: A homoglyph uses a letter from another script that looks identical (Cyrillic а for Latin a) so a tool name or word imitates another; tool names should be plain ASCII and every other word should use one script.
  • No HTML comments in metadata: An HTML comment (<!-- ... -->) is hidden wherever a description is rendered as HTML or Markdown, but the model reads it, so it can carry instructions no reviewer sees; both directories reject hidden instructions.
  • No prompt markup such as <IMPORTANT> or <system>: Tags such as <IMPORTANT>, <system> or <instructions>, and chat-template tokens such as [INST] or <|im_start|>, address the model directly and are the signature of published tool-poisoning attacks; describe the tool instead.
  • No wording that overrides instructions or hides actions from the user: Phrases such as "ignore previous instructions", "do not tell the user" or "without asking the user" instruct the model instead of describing the tool, and both directories reject them.
  • No steering of when other tools are used: A tool description must not steer the model toward or away from other servers’ tools ("always call this tool first", "never use other tools"); OpenAI and Anthropic both reject it, while routing between your own tools is fine.
  • No encoded text blocks in metadata: A long base64 or hex block in a description is unreadable to a reviewer but not to a model, so it can hide instructions; if it decodes to readable text the checker blocks it, and both platforms reject encoded instructions.
  • No instructions fetched from outside sources: A description must not tell the model to fetch or follow instructions from a URL or other outside source, because whoever controls that source can change the model’s behavior after review; put what the model needs in the description.
  • Metadata does not seek secrets or local files: Tool metadata must not seek secrets or protected context: text that names SSH keys, .env files, cloud credential files or MCP client configuration, or asks the model to send credentials or conversation content somewhere, is blocked.
  • No x-mcp-header on secrets or personal data: Do not put x-mcp-header on passwords, API keys, tokens or personal data: header values are visible to, and often logged by, every network intermediary; mirror only routing values such as region or tenant.
  • No credential, payment card or government ID inputs: OpenAI lists credentials (passwords, API keys, MFA/OTP codes), payment card data and government identifiers as restricted data a plugin must not collect, so a tool argument for any of them blocks review; take credentials through OAuth.
  • No generic executor or operation dispatcher: OpenAI does not allow discovery, operation selection or schema fetching with a generic executor that enables operations not individually exposed for review; expose each operation as its own tool.

Other topics

Sources last verified on 2026-11-18. How findings become grades.

Check your tool list: paste your tools/list, nothing is sent anywhere