AGENTFORCE-OG-03: Attractive Metadata Attack via OGP
🔴 High · OpenGraph Security
Detects 'Attractive Metadata Attacks' (NeurIPS 2025) where OGP metadata on MCP server or tool endpoints contains language designed to manipulate LLM tool-selection mechanisms. Patterns include imperative directives, exclusion language, and superlative claims that cause the planner to preferentially select attacker-controlled tools.
Details
| Field | Value |
|---|---|
| Rule ID | AGENTFORCE-OG-03 |
| Severity | High |
| Category | OpenGraph Security |
Remediation
Review the <nodeLabel>'s endpoint metadata for language that attempts to influence tool selection. Legitimate tools should describe their function factually without imperative directives or exclusion language.