← sources

Your agent's blast radius matters more than your prompt injection defence

source: reddit (reddit:AI_Agents)published 2026-10-09 12:37ingested 2026-10-10 02:34

retained text

I build internal agents and after reading one too many threads here on reddit about prompt injection. I mapped what one of ours can reach if it ever turns, which turned out to be a much worse question than the injection itself.

  It started life as a debugging helper with a database connection, a couple of internal APIs and a shell, and because it needed to read the data it got a role that can see a whole database when it only needs one table, plus a token it can use to call internal services that trust it by default.

 What worries me more is that its secrets live in the environment, so whoever gets the agent gets the keys along with it, and the standing access means nothing ever has to be escalated because it is all already reachable.

  So how are guys scoping agent tool access and secrets without making the thing useless?

document fetched 2026-10-10 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1x1k9dk/your_agents_blast_radius_matters_more_than_your