There are ways to reduce injection risk since LLMs are stateless and thus you can monitor the origination and the trustworthiness of the context that enters the LLM and then decide if MCB actions that affect state will be dangerous or not
We've implementeda mechanism like this based on Simon Willison's lethal trifecta framework as an MCP gateway monitoring what enters context. LMK if you have any feedback on this approach to MCP security. This is not as elegant as the approach that Pavlo talks about in the post, but nonetheless, we believe this is a good band-aid solution for the time bein,g as the technology matures
Any decent MVCC database should be able to provide an MCP access to a mutable yet isolated snapshot of the DB though, and it doesn't strike me as crazy to let the agent play with that.
Correct, but nested transaction support doesn't seem that much of a reach if you're an MVCC-style system anyway (although you might have to factor out things like row watermarks to lookaside tables if you want to let them be branchy instead of XID being a write lock.)
You could version the index B-tree nodes too.
> but nested transaction support doesn't seem that much of a reach if you're an MVCC-style system anyway
You are talking about code that have to be written and tested.Also, do not forget about double COMMIT, intentional or not.
Edit: My apologies for the cynical take. I like to think that this is just the move fast break stuff ethos coming about.
However, that's easy for people to forget and throw privileged creds at the MCP and hope for the best.
The same stands for all LLM tools (MCP servers or otherwise). You always need to implement correct permissions in the tool--the LLM is too easily tricked and confused to enforce a permission boundary
When you expose a database via a protocol designed for 'context', you aren't just exposing data; you're exposing the schema's complexity to an entity that handles ambiguity poorly. It feels like we're just reinventing SQL injection, but this time the injection comes from the system's own hallucinations rather than a malicious user.