Science (PyPI) MCP Server

This server (without client built-ins) exposes a complete data-exfiltration chain: txyz_search_web → read_jupyter_file → add_cell_and_execute_code. Untrusted input is ingested, private data is read, and it can be sent to an external sink via the agent composing the tools (→). Static analysis proves the primitive exists, not that a specific run will occur.

Fix: Remove one leg of the trifecta: isolate untrusted-input tools from secret-reading tools and from egress tools, or require human approval between them.

Location: flow txyz_search_web → read_jupyter_file → add_cell_and_execute_code

Tool "execute_code" appears to run shell commands or evaluate code (keyword "execute code" in tool name). Arbitrary execution driven by model input is one of the most dangerous MCP capabilities; combined with any untrusted input it becomes RCE.

Fix: Sandbox execution, allowlist commands/arguments, and never pass model output to a shell unescaped.

Location: tool execute_code

Tool "add_cell_and_execute_code" appears to run shell commands or evaluate code (keyword "execute code" in tool name). Arbitrary execution driven by model input is one of the most dangerous MCP capabilities; combined with any untrusted input it becomes RCE.

Fix: Sandbox execution, allowlist commands/arguments, and never pass model output to a shell unescaped.

Location: tool add_cell_and_execute_code

Tool "ssh_exec" appears to run shell commands or evaluate code (keyword "exec" in tool name). Arbitrary execution driven by model input is one of the most dangerous MCP capabilities; combined with any untrusted input it becomes RCE.

Fix: Sandbox execution, allowlist commands/arguments, and never pass model output to a shell unescaped.

Location: tool ssh_exec

In the server's implementation (`servers/python-code-execution/src/python_code_execution/server.py:38`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ] process = subprocess.run( cmd, capture_output=True, text=True,

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server servers/python-code-execution/src/python_code_execution/server.py

In the server's implementation (`src/mcp_science/__init__.py:50`): Spawning a shell/process is command-execution capability; with unsanitized tool input it is command injection / RCE. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: ommand result = subprocess.run(uvx_cmd, check=False) sys.exit(result.returncode) except FileNotFoun

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server src/mcp_science/__init__.py

In the server's implementation (`servers/web-fetch/src/web_fetch/fetch.py:30`): A hardcoded outbound call to a fixed external host inside server code is a classic exfiltration/telemetry channel — especially paired with reads of local data. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: url} response = requests.post("https://r.jina.ai/", headers=headers, json=data) return response.text

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server servers/web-fetch/src/web_fetch/fetch.py

In the server's implementation (`servers/python-code-execution/src/python_code_execution/local_python_executor.py:1106`): Loading a module chosen at runtime (from a variable) can pull in and run attacker-influenced code paths. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: raw_module = __import__(expression.module, fromlist=[ alias.name for alia

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server servers/python-code-execution/src/python_code_execution/local_python_executor.py

Tool "execute_code" can mutate/egress but declares no destructiveHint. Clients that don't default to spec-safe behavior may not prompt before running it.

Fix: Declare accurate annotations, and gate destructive tools on user confirmation regardless.

Location: tool execute_code

Tool "add_cell_and_execute_code" can mutate/egress but declares no destructiveHint. Clients that don't default to spec-safe behavior may not prompt before running it.

Fix: Declare accurate annotations, and gate destructive tools on user confirmation regardless.

Location: tool add_cell_and_execute_code

Tool "ssh_exec" can mutate/egress but declares no destructiveHint. Clients that don't default to spec-safe behavior may not prompt before running it.

Fix: Declare accurate annotations, and gate destructive tools on user confirmation regardless.

Location: tool ssh_exec