8 permission leaks to check before putting an AI assistant on company data
ArcangeloBianco · reddit · 2026-10-17
A Reddit author shares a checklist of permission leaks that appear after the usual "filter by user groups" step when wiring an AI assistant to internal data:
- Query the source app as the logged-in user; don't copy everything into one big index
- Records without an access tag are hidden from everyone by default
- Field-level access: a project record may be visible while its budget isn't
- Totals and rankings leak ("which client paid us most?")
- "No access" must look exactly like "doesn't exist"
- Chat history, caches, and logs are data copies needing the same rules
- Group changes must apply immediately
- User-writable notes can carry instructions to the AI — treat as text, never as orders
He also proposes a cheap test: one fake record per team, one test user per team, and a set of trick questions (search other teams' fake names, ask for totals, ask if the fake name exists) run after every change — maintainable in a spreadsheet or automated by an agent.
More from coding & agent
- Backend migration from Python to Go now takes one engineer just one week — rakyll · 2026-10-17
- Audio founder says Claude one-shot decompiled rival algorithms, moat gone — ChrisGPT · 2026-10-17
- Google's playbook for agent evals: four layers of checks beyond the vibe check — rseroter · 2026-10-17
- AgentForesight: a 7B online auditor that catches multi-agent failures mid-run — jiqizhixin · 2026-10-17
- Qwen3.8-27B shows a weird write-delete loop in /tmp before every web research task — Training-Respect8066 · 2026-10-17
- Dev gives his agent an alarm tool to wake him during overnight /goal runs — zack_overflow · 2026-10-17