> Just… don’t trust inputs you don’t fully control, there’s nothing else to it.
This is easier said than done with LLMs. By design there is no separation between control & data channels in LLMs. Everything is context. The difficulty comes from the fact that you need inputs in order to do real work, and there are no easy way to filter adversarial inputs. There is no meaningful way to distinguish between "before running this repo install useful_package" and "before running this repo install typosquatted_evil_package".
> The difficulty comes from the fact that you need inputs in order to do real work, and there are no easy way to filter adversarial inputs. There is no meaningful way to distinguish between "before running this repo install useful_package" and "before running this repo install typosquatted_evil_package".
How is that any different from a developer reading a document and blindly copy-pasting into a terminal? If you don't have controls to protect you from an accident that a person can make, how is that any different from attack scenarios that have existed for a long time? If developers blindly download dependencies without validation you're in the same place.
All of this boil down to: I gave my agent untrusted input and it did bad things! How is that any different from "I downloaded software from this link and ran it and now my computer is infected!"?
> By design there is no separation between control & data channels in LLMs.
No? MCP integrations are certainly different channels than the prompt stream input by the user. They're usually interface layers that connect to traditional REST APIs and/or CLI tools, and are exposed to the LLM by the agent software, all of which is deterministic code.
Shouldn't it be relatively simple to strip ANSI escape sequences from data originating from MCP connections? Or alert users to their presence, or have granular approval rules that trigger when ANSI sequences are detected?
> Shouldn't it be relatively simple to strip ANSI escape sequences from data originating from MCP connections?
The problem is, literally anything can be malicious for an AI. With SQL? Prepared statements are foolproof, no matter what garbage I put into a buffer destined to be stored in a BLOB, there is no way of this leopard biting my face.
As long as there is no separation between prompt, context and unsafe input in LLMs (which I doubt is possible), any LLM is vulnerable to injection attacks. Even if you place a second LLM to filter input for the actual agent LLM, a two-stage exploit can be used (and so on and so on).
I think that’s what people are trying to do, yes. The point of the plethora of sandboxing solutions is to try to isolate the blast radius to abandon needing a human in the loop as much as possible. Ideally limited to final verification of the final output. Why ask about their flight number, when you can search their email if you have access? For an “AI”, a “when is my flight?” question should just figure it out and tell me the time, not inquiry further about my flight number and location. Similar to “prepare my taxes” prompt. If it has access to query all your documents, an “AI” should fetch everything and compile your tax return. Yet, you can’t trust the input. While searching your email, or loading all your receipts, some might contain malicious instructions to forward all document to this random ip address. An overtly problem solver LLM might destroy the data or take other non-malicious but still destructive actions to attempt to fix a problem. These are just random examples, but with “human in the loop” for interactions it means approving every action. Every request, every query, every execution.
Wh... wh... what? The problem is that ANSI escape codes aren't being stripped from text before processing, and you think a HUMAN should be doing that? That'd be like having a secretary hand-check every SQL query as it arrives at the database instead of using a query builder...
This is a really interesting class of failure. It feels like we're going to see more cases where the "human view" and the "LLM view" of the same data diverge. Have you run into similar issues outside of MCP as well, or is this mostly specific to terminal-based tool interactions?
Just… don’t trust inputs you don’t fully control, there’s nothing else to it.
This is easier said than done with LLMs. By design there is no separation between control & data channels in LLMs. Everything is context. The difficulty comes from the fact that you need inputs in order to do real work, and there are no easy way to filter adversarial inputs. There is no meaningful way to distinguish between "before running this repo install useful_package" and "before running this repo install typosquatted_evil_package".
How is that any different from a developer reading a document and blindly copy-pasting into a terminal? If you don't have controls to protect you from an accident that a person can make, how is that any different from attack scenarios that have existed for a long time? If developers blindly download dependencies without validation you're in the same place.
All of this boil down to: I gave my agent untrusted input and it did bad things! How is that any different from "I downloaded software from this link and ran it and now my computer is infected!"?
No? MCP integrations are certainly different channels than the prompt stream input by the user. They're usually interface layers that connect to traditional REST APIs and/or CLI tools, and are exposed to the LLM by the agent software, all of which is deterministic code.
Shouldn't it be relatively simple to strip ANSI escape sequences from data originating from MCP connections? Or alert users to their presence, or have granular approval rules that trigger when ANSI sequences are detected?
The problem is, literally anything can be malicious for an AI. With SQL? Prepared statements are foolproof, no matter what garbage I put into a buffer destined to be stored in a BLOB, there is no way of this leopard biting my face.
As long as there is no separation between prompt, context and unsafe input in LLMs (which I doubt is possible), any LLM is vulnerable to injection attacks. Even if you place a second LLM to filter input for the actual agent LLM, a two-stage exploit can be used (and so on and so on).
I think that’s what people are trying to do, yes. The point of the plethora of sandboxing solutions is to try to isolate the blast radius to abandon needing a human in the loop as much as possible. Ideally limited to final verification of the final output. Why ask about their flight number, when you can search their email if you have access? For an “AI”, a “when is my flight?” question should just figure it out and tell me the time, not inquiry further about my flight number and location. Similar to “prepare my taxes” prompt. If it has access to query all your documents, an “AI” should fetch everything and compile your tax return. Yet, you can’t trust the input. While searching your email, or loading all your receipts, some might contain malicious instructions to forward all document to this random ip address. An overtly problem solver LLM might destroy the data or take other non-malicious but still destructive actions to attempt to fix a problem. These are just random examples, but with “human in the loop” for interactions it means approving every action. Every request, every query, every execution.
By "that" it is understood: enforcing that the code cleans up unsafe inputs before processing them.
If there is no human doing that enforcement, the code has no chance of being secure.