LLM-as-Untrusted-Component

Description

Architectural stance emerging from the speaker attribution bug: even when an LLM is acting as an agent on behalf of a user, its turn metadata and self-generated messages must be treated as untrusted input. Prompt engineering and permission boundaries do not fix a system that cannot reliably track who said what; the fix is architectural (cryptographic turn signing), not behavioral.

Key claims

Relations

Sources

src-20260419-3e34d5830692