🧪 Active Investigation

Function Call Leakage

Function call leakage happens because AI hallucinates tool usage, and may be fixable via better prompts or MCP metadata.

CURRENT POSITION

I believe function call leakage is mainly caused by insufficient prompt or metadata constraints around tool calling, and MCP-level fixes might reduce hallucinated calls.

KEY ASSUMPTIONS

  • AI models hallucinate tool calls when tool boundaries are not clearly enforced.
  • Better prompts or metadata in MCP can reduce or prevent function call leakage.

SUPPORTING EVIDENCE

  • Personal experience observing AI claiming it called tools when it did not.

OPEN QUESTIONS

  • Whether prompt engineering alone is sufficient or if architectural changes are required.
  • What specific MCP metadata patterns are most effective at preventing leakage.

WANT TO EXPLORE DEEPER?

Read Full Thought →

by parag