Shared MCP server credentials leak data across users — three fixes and their tradeoffs
AuthByExample · reddit · 2026-10-18
A common setup — one internal MCP server per system holding a single service credential — lets any user's agent read data their app permissions would block. Per-tool allowlists don't help when the leak is inside an allowed tool like searchtickets running as the service account. The author lays out three fixes: passing user OAuth tokens downstream (clean but often unsupported), gateway-level per-resource checks (slow, duplicates permission logic), and pushing identity into queries via Postgres row-level security. Asks how others handle list-returning tools.
More from coding & agent
- DBeaver: free open-source DB tool with built-in AI chat, 100+ databases — alex_verem · 2026-10-18
- Indie hacker makes $8,000 in two weeks with SEO, releases OpenSEO × Claude Code tutorial — yihui_indie · 2026-10-18
- How developers are handling long-term agent memory: Mem0, Zep, or DIY — Efficient_Joke3384 · 2026-10-18
- Indie maker's fully AI-agent-run YouTube channel notches first video past 1K views — tibo_maker · 2026-10-18
- Luma adds Send to Luma inside Claude Motion, turning animated explainers into real video projects — lmoroney · 2026-10-18
- Small businesses don't need agents: define the workflow first, pick the tool second — DavidLinthicum · 2026-10-18