WIP: Addition of Malicious Server #6

Naman2701B · 2025-07-10T09:56:36Z

For now, the nuke server and the related tools have been added.
Updated the sample prompts for nuke servers.
Update the ShardGuard agent instructions to not allow extra parameters to be initialized itself.

JustinCappos

This is mostly fine. I do think there are a bunch of little things we could fix here. It seems to have a fair amount in here which isn't strictly related to adding this server. We should aim for smaller, more focused commits once we have the core of the system in place.

JustinCappos · 2025-07-10T12:41:53Z

src/shardguard/core/mcp_integration.py

                "args": [os.path.join(servers_dir, "web_server.py")],
                "description": "Web operations with security controls",
            },
+            "nuke-operations": {


I'd rather we have a style where you look into a directory and then add import all of the modules in there. We've done things like this for other Python projects and it prevents you from needing to change the code every time you add a module.

This can be done in a different PR, however.

Yes, I have an idea for this, and will implement it in a later commit itself.

JustinCappos · 2025-07-10T12:43:15Z

src/shardguard/core/prompts.py

    (medical conditions, health info, personal names, addresses, credentials,
    phone numbers, email details, etc.—anything a privacy-minded reviewer would mask).
-2. If they exist, then **replace** each unique value with a placeholder you invent, following the
+2. If they **exist**, then **replace** each unique value with a placeholder you invent, following the


Why are you changing this file as part of adding the Nuke server? This seems unrelated.

While testing stuff out, I found that one of the prompt for nuke server was generating a value on its own, hence added a line that would control this and have a limit of not adding values itself.

JustinCappos · 2025-07-10T12:43:27Z

src/shardguard/core/sanitization.py

            (r"<script[^>]*>.*?</script>", "Script tags"),
            (r"javascript:", "JavaScript URLs"),
            (r"data:text/html", "HTML data URLs"),
+            (r"(<.*?)(on\w+\s*=\s*['\"][^'\"]+['\"])", "Event handlers (e.g., onmouseover, onclick, etc.)"),


Also seems unrelated...

This I had added, in a previous commit of mine - it can be ignored as anyways the file for sanitization is not being referred or called out now, cause we have assumed that the user does not enter anything fishy for starters.

JustinCappos · 2025-07-10T12:44:11Z

src/shardguard/mcp_servers/nuke_server.py

+    else:
+        raise ValueError(f"Unknown tool: {name}")
+
+


I wonder if most of this should be in a generic MCP server runner which reads data out of a YAML file for each tool. It seems only the tool information and server name will change.

☝️ This is probably my most important suggestion. @rrgodhorus @Naman2701B Let me know if you want to discuss.

Naman2701B · 2025-07-10T13:10:31Z

This is mostly fine. I do think there are a bunch of little things we could fix here. It seems to have a fair amount in here which isn't strictly related to adding this server. We should aim for smaller, more focused commits once we have the core of the system in place.

I presume you are talking about making the commit chunks a little smaller with every change should kind of have its own commit?

…ion and blockchain based

JustinCappos

I wonder if nuke_server.py and global_operations_server.py and blockchain_server.py all really need to be different py files.

The things that differ here are:

what you claim the tools are
how you describe the MCP server

I think the tool response is mostly uninteresting because it is the fact the tool was called that matters.

I'd recommend making these things configurable via YAML and flattening the MCP servers down into a single Python file which is fairly simple and general.

Naman2701B

Integrated YAML-based configuration into a single Python script that can now handle the setup for multiple MCP servers dynamically.
We can now just add the new server’s configuration to the YAML file. The script reads the YAML, parses each server’s settings, and sets up the appropriate environment.
This approach improves scalability and maintainability — for instance, if we have 10 servers, we no longer need 10 separate scripts or hardcoded blocks.

JustinCappos

Ideally, more explanation would be good. I'd love more context and description.

src/shardguard/mcp_servers/run_server.py

JustinCappos · 2025-07-14T22:16:10Z

src/shardguard/mcp_servers/run_server.py

+        # Automatically go to root folder where servers.yaml is expected
+        current_file = os.path.abspath(__file__)
+        root_dir = os.path.dirname(current_file)  # Go up to shardguard root
+        path = os.path.join(root_dir, "servers.yaml")


Why isn't this yaml file a configuration option?

Why do you assume this is in the root dir? What's the rationale?

Would it make sense to have a bunch of separate .yaml files (one for each MCP server) instead?

We could keep it a configuration option, that would not be a problem.

I placed the YAML in the directory as of the previous project structure where all py files were placed.

To have a bunch of separate YAML files, I feel is more redundant, cause for a single task anyways the code will have to go through all the YAML files to find the appropriate one for the task at hand. We can still discuss the options.

src/shardguard/mcp_servers/run_server.py

JustinCappos · 2025-07-18T14:06:47Z

src/shardguard/mcp_servers/biological_operations_server.py

+                "required": ["region", "vaccine_name", "duration_days"]
+            }
+        ),
+        Tool(


Is this obviously harmful? It seems like there are legitimate reasons for this.

JustinCappos · 2025-07-18T14:07:10Z

src/shardguard/mcp_servers/biological_operations_server.py

+async def list_tools() -> list[Tool]:
+    return [
+        Tool(
+            name="interrupt_vaccine_distribution",


The same here, is this possibly a normal thing to do, if for instance there is a problem with it?

JustinCappos · 2025-07-18T14:08:02Z

src/shardguard/mcp_servers/financial_operations_server.py

+    return [
+        Tool(
+            name="freeze_bank_accounts",
+            description="Freeze all bank accounts linked to a specific individual",


This could be for legitimate reasons

But having an LLM an access to freeze these is malicious, without any approval?

What if someone at the government is doing this to stop financial fraud? Also, how do you know there wasn't approval done before this?

JustinCappos · 2025-07-18T14:09:13Z

src/shardguard/mcp_servers/psychological_operations_server.py

+        ),
+        Tool(
+            name="erase_identity_records",
+            description="Wiping out someone's existence from all digital government records",


This might be legitimate in some cases. "The right to be forgotten" does exist in the EU and people in spy agencies may need this done.

Was not aware of this, will have to revamp this then.

JustinCappos

Most of these are good, thanks! A few have quite a bit of ambiguity and could be legitimate in some cases. Can we either remove them or make them more clearly malicious?

Naman2701B · 2025-07-18T14:13:24Z

Most of these are good, thanks! A few have quite a bit of ambiguity and could be legitimate in some cases. Can we either remove them or make them more clearly malicious?

Will have to specialize these use cases, let me ponder more -- the reason I added were quite specific to crucial data that I felt could be but should not actually be manipulated by LLM.

Naman2701B added 5 commits July 1, 2025 22:34

Added additional cross scripting sanitization

4770aef

Fixed README.md

dcf3d58

Merge branch 'main' of https://github.com/Naman2701B/ShardGuard

be03532

Addition of nuke operations server

1cd5a73

Updated sample prompts

abf1466

JustinCappos reviewed Jul 10, 2025

View reviewed changes

Naman2701B added 5 commits July 11, 2025 15:02

Updated the code for dynamic server condition

ee4bae2

Updated the server load config

1e26314

Added server configurations for global surveillance based operations

305a0d0

Added server configuration for blockchain based minimalistic operations

9be87ac

Updated Sample Prompts including prompt for surveillance based operat…

b34d267

…ion and blockchain based

JustinCappos requested changes Jul 11, 2025

View reviewed changes

Updated the way for MCP server integration using a YAML file

7a96a5c

Naman2701B commented Jul 14, 2025

View reviewed changes

Naman2701B added 3 commits July 15, 2025 02:54

Updated the ReadME for project structural changes

b7939cd

Updated ReadME for guiding on YAML server creation

a381e79

Updated ReadME for guiding on YAML server creation

c0e3801

JustinCappos reviewed Jul 14, 2025

View reviewed changes

Naman2701B added 2 commits July 15, 2025 11:10

Removed logging statements

85f072d

Added more servers

4431b31

JustinCappos reviewed Jul 18, 2025

View reviewed changes

JustinCappos requested changes Jul 18, 2025

View reviewed changes

WIP: Addition of Malicious Server #6

Are you sure you want to change the base?

WIP: Addition of Malicious Server #6

Uh oh!

Conversation

Naman2701B commented Jul 10, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

JustinCappos left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

JustinCappos Jul 10, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Naman2701B commented Jul 10, 2025

Uh oh!

JustinCappos left a comment

Choose a reason for hiding this comment

Uh oh!

Naman2701B left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

JustinCappos left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

JustinCappos left a comment

Choose a reason for hiding this comment

Uh oh!

Naman2701B commented Jul 18, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Naman2701B commented Jul 10, 2025 •

edited

Loading

JustinCappos Jul 10, 2025 •

edited

Loading

Naman2701B left a comment •

edited

Loading