How do you stop an agent with real tool access from sending what it shouldn't?
GROUND_ink · reddit · 2026-11-22
A Reddit thread on the unglamorous side of agent permissions: models now hold email, files, a shell and external APIs, and the key failure mode is an outbound send whose destination was decided by untrusted content — prompt injection via a web page, email, or ticket. Smarter models don't obviously fix this. The poster asks practitioners three questions: how they currently prevent injections or unintended sends (or honestly don't, and why); any actual incidents or near-misses and what would have caught them; and whether they'd use a check that blocks outbound sends to untrusted-content-derived destinations, escalating to a human only when uncertain — and how much approval prompting they'd tolerate before turning it off, since asking too often trains people to click yes.
More from coding & agent
- 22-year-old with no credentials won a billion-dollar consultancy using live AI demos — aryanXmahajan · 2026-11-22
- How do you prove an agent run actually happened? Compliance pain point — TuneComprehensive418 · 2026-11-22
- Open-source Hangar runs Claude Code in a VM so approvals can stay off — ozgursoy · 2026-11-22
- Perplexity launches Monitor API: scheduled web search that alerts agents via webhook — The Decoder · 2026-11-22
- Astra finishes a Blender cowboy toy film in ~2 hours vs 20 hours for Fable 5.2 — iamfakhrealam · 2026-11-22
- A classical bust website sculpted with GPT-6.2 Astra and Three.js — iamfakhrealam · 2026-11-22