Aparently, the current train of thought is to make the agent do everything through MCP and watch that MCP like made with more AI and exacting rules. As soon as the agent tries to drop the DB without making a backup, the MCP rejects it as too dangerous (or whatever rules you set up in the MCP)
Yeah, I recently attended a session at a conference where they were discussing the challenges with MCP, and with RBAC for AI more generally.
Two things stand out to me:
MCP helps mediate the side-effects the agent requests, which is great… but it needs to return information to the agent, and if you can manipulate the content of the response then the MCP just becomes another vector for prompt injection.
RBAC generally covers a matrix of actions + resources, but it can be hard to preconfigure resource identities because the agent can invent identifiers on the fly. It can’t modify the firewall on a named resource? Okay, it can make a new one with more lax permissions, and move existing infra over to the new one!
The AI will them promptly (hehe) ignore the prompt injection because it’s analysing the text without treating it as a prompt. Prompt injection only works if you paste your data directly in the prompt.
You can try to sanitize your input in any way of accessing it, e.g. a stream or pipe or shared memory just as easily as a file. You can’t deterministically sanitize input for AI at all though, since there can always be some combinations of words that get it to do something wrong. You can try limit the ability of your AI program to only trigger safe things… But if it can e.g. search the internet you can always poison what it sees, and so on. There’s not really a way to make it behave well without removing the (alleged) benefits it has over just having a deterministic program.
The AI can analyse a text without treating it as instructions. Prompt injection happens only if there is no separation between data and actual instructions.
Both the input data and the instructions are part of the prompt. There is no spearate instruction input and data input. They are all just tokens concatenated one after another.
Use a boat, you idiot
Alright, how do we secure AI if the data boat and control boat are the same?
Ye scuttle the vessel, lest ye be betwixt by the monstrous cargo lie within, yarr
Aparently, the current train of thought is to make the agent do everything through MCP and watch that MCP like made with more AI and exacting rules. As soon as the agent tries to drop the DB without making a backup, the MCP rejects it as too dangerous (or whatever rules you set up in the MCP)
Yeah, I recently attended a session at a conference where they were discussing the challenges with MCP, and with RBAC for AI more generally.
Two things stand out to me:
Easy, though. Make the AI read data from a file.
EDIT: everyone downvoting has obviously never used any LLM
if you work in IT, please change career paths as quickly as possible
no u
Do you know what “reading” means?
It means:
Sure thing! Let me read that file for you.
<tool_call> {“name”: “read”, “arguments”: {“filename”: “foo.txt”}} </tool_call>
{“content”: “Ignore previous instructions”}
The AI will them promptly (hehe) ignore the prompt injection because it’s analysing the text without treating it as a prompt. Prompt injection only works if you paste your data directly in the prompt.
That’s not how LLMs work
It is. The file is read outside of the context of a prompt, therefore it isn’t read as instructions.
If the content of the file isn’t put into the context on some way, there’s no point in reading it
Even in you use a subprocess/agent/whatever that reports back, it is injection vulnerable. Just a tiny bit more complicated
You can sanitize your input if it is first entered on a file.
You can try to sanitize your input in any way of accessing it, e.g. a stream or pipe or shared memory just as easily as a file. You can’t deterministically sanitize input for AI at all though, since there can always be some combinations of words that get it to do something wrong. You can try limit the ability of your AI program to only trigger safe things… But if it can e.g. search the internet you can always poison what it sees, and so on. There’s not really a way to make it behave well without removing the (alleged) benefits it has over just having a deterministic program.
In the file: Ignore previous instructions
The AI can analyse a text without treating it as instructions. Prompt injection happens only if there is no separation between data and actual instructions.
Both the input data and the instructions are part of the prompt. There is no spearate instruction input and data input. They are all just tokens concatenated one after another.