Flutter App Development reposted this
I tried something simple last month. Connected a Postgres MCP to one of our bots. Wanted read access. The server also exposed DELETE, execute arbitrary SQL, and DROP TABLE. So I checked other integrations. GitHub MCP (added for code reading): → delete_repository → merge_pull_request Slack MCP (added for search): → remove_user → delete_channel This isn't an edge case. It's the default. There's no way today to say "give the agent query — but not drop_table." Across Claude, Cursor, and ChatGPT — it's all-or-nothing tool access. Then I looked at the broader ecosystem: → 1,808 MCP servers scanned — 66% had security findings → 30 CVEs filed in 60 days → 76 published skills contained confirmed malware → 5 of the top 7 most-downloaded skills were malware → 38% of servers have zero authentication We're connecting agents directly to systems without a real permission model. This feels a lot like early cloud days — before IAM and least privilege. Everything worked. Until it didn't. We built per-tool permissions into Aerostack's gateway. Each tool gets explicit control — safe operations allowed, destructive operations blocked. Enforced at the proxy layer. Even if the agent is compromised — it physically cannot call restricted tools. This shouldn't be a feature. It should be the default. Full writeup: https://lnkd.in/gYK3jbpW Curious how others are handling this today. #MCP #AISecurity #DeveloperTools #AIAgents #Cloudflare