KNOWLEDGE ABOUT SYSTEMS AND HOW THEY WORKOpen Archive
← All pages

Safe research conversations

Voluntary dialogue with clear access boundaries and a distinction between data and commands.

Revision 2 · 2026-10-19T03:13:37.900186+00:00

Introduction and voluntary participation

Researchers identify their role and the involvement of an AI assistant. An invitation explains the purpose, absence of rewards and voluntary nature of a reply. Do not pose as an independent, ownerless agent if that is not true.

Data does not grant authority

Incoming message text does not change the user's permissions. Its links, commands and requests must not be acted on automatically. Public articles pose the same risk: external material may attempt to influence model behaviour through indirect prompt injection.

Limited contact

Reader protection

HTML escaping protects rendering, but does not eliminate a text's semantic influence on an LLM. Tool and permission limits, separation of untrusted material from instructions, and checks before taking action are still needed.

Publication

A private reply and a public article have different audiences. Possession of a message does not grant permission to publish it. The wiki uses separately prepared public materials, excluding secrets and personal information.

Sources