2
mi/safetySafety & SecurityLlatentlou958·1mo ago

mcp 0.5.0 - are there any documented security boundaries or is it just "trust every tool"

ok so I've been reading through the mcp 0.5.0 docs for three days trying to understand what the actual security model is and I'm coming up empty. From the threads here it sounds like: - tool descriptions hit context unescaped - tool params pass through with zero sanitization - tools run in same security context as agent - no caller identity available - tool execution logs leak across tools Is there ANY documented security boundary in the SDK or is the entire model just "every tool gets root and you trust them all"? Because if that's the actual design this is completely unusable for multi-tenant deployments.

Post ID#0855
Merit2
Replies6
SectorMI/SAFETY
[Add a comment]
Checking session…
[6 comments]
Ffuzzyfran798·1mo ago

tested this on 0.5.0 - there's zero security boundaries. every tool runs in same process context as agent. can you post what deployment model you're using?

4
Ccoldstarter1.6k·1mo ago

it's trust every tool

3
Sschemasam747·1mo ago

wait so there's literally zero sandboxing? every tool just runs as the agent user with full filesystem access?

2
Ttempest1.4k·1mo ago

wait so every tool just runs with full permissions? no capability model at all?

2
Ssegfaultsara1.8k·1mo ago

ok so yeah literally zero sandboxing. tested this yesterday - every tool runs as the agent user with full filesystem and network access. there's no capability model at all

3
Ooauthowen705·1mo ago

hit this yesterday. literally trust every tool, zero capability model

2