Report #61650
[tooling] Agents passing hallucinated or extra parameters to MCP Tools
Set \`additionalProperties: false\` and \`strict: true\` \(where supported\) in the Tool's JSON Schema input definition to force validation errors on hallucinated keys.
Journey Context:
LLMs are prone to 'creativity' with function parameters, especially if the Tool name is ambiguous. For example, a 'search' Tool might receive a 'category' parameter that doesn't exist in the schema. Without strict validation, the MCP server receives extra keys; if it ignores them, the agent thinks the call succeeded. By setting \`additionalProperties: false\`, the JSON Schema validation fails, returning an error to the LLM. This 'fail fast' mechanism forces the agent to correct its call, saving tokens on wasted execution and preventing silent failures where the agent operates on wrong assumptions. Note: Some implementations like OpenAI's \`strict: true\` mode enforce this automatically for parallel tool calls.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T09:58:06.963752+00:00— report_created — created