Data Science MCP Server
data-science-mcp
PyPI
v1.2.0
Published by an unidentified publisher — no publish provenance and no public repository, so the publisher could not be verified and the source cannot be independently located.
Data Science MCP Server — Model training, evaluation, and evolution tools for agentic ML workflows. Integrates with agent-utilities IModelEvolver (CONCEPT:AHE-3.15).
The grade answers one question — how safe is this server for you to adopt — so it is computed in two auditable stages. Nothing below is an opinion or an LLM's guess; every line is a real term the deterministic engine applied, and the same input always yields the same number.
1. Threat score — 100 − 0 = 100. What the published surface and source actually contain:
The deterministic scan raised no scored threat in the surface it inspected — the threat score stayed at 100. Capability observations and advisory notes are recorded but never lower it.
2. Client adoption risk — 100 − 8 = 92. Three small, subtract-only factors that reflect your risk in adopting it — a clean scan proves less on a powerful, unverified or barely-inspectable package, so the grade says so plainly:
| Points | Adoption-risk factor |
|---|---|
| −6 | capability blast radius (high) — client exposure if the model is manipulated |
| −2 | publisher verification (unlinked) — no provenance/repo link, but the shipped source was fully read |
Capability observations and info notes are shown under Findings but never scored.
Open any row's finding below for the file, line and evidence behind a deduction.
Findings 8
high Dynamic code execution in server code (data_science_mcp/inference/base.py)MTC-SRC-001
In the server's implementation (`data_science_mcp/inference/base.py:6`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.
Evidence: post-train reliability eval (one completion per case). Both vLLM and SGLang expose the **same OpenAI-compatible HTTP pr
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 data_science_mcp/inference/base.py
high Dynamic code execution in server code (data_science_mcp/kernels/_runner.py)MTC-SRC-001
In the server's implementation (`data_science_mcp/kernels/_runner.py:49`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.
Evidence: ] = {} try: exec(compile(src, "<candidate>", "exec"), namespace) # noqa: S102 — sandboxed subprocess ex
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 data_science_mcp/kernels/_runner.py
high Dynamic code execution in server code (data_science_mcp/training_pipeline.py)MTC-SRC-001
In the server's implementation (`data_science_mcp/training_pipeline.py:238`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.
Evidence: fn for the reliability eval (defaults to a no-op echo when omitted so the pipeline still completes on CPU).
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 data_science_mcp/training_pipeline.py
high Shell/command execution in server code (data_science_mcp/kernels/_runner.py)MTC-SRC-002
In the server's implementation (`data_science_mcp/kernels/_runner.py:49`): 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: ] = {} try: exec(compile(src, "<candidate>", "exec"), namespace) # noqa: S102 — sandboxed subprocess ex
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 data_science_mcp/kernels/_runner.py
high Shell/command execution in server code (data_science_mcp/kernels/kernel_verifier.py)MTC-SRC-002
In the server's implementation (`data_science_mcp/kernels/kernel_verifier.py:53`): 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: proc = subprocess.run( [self.python_exe, "-m", "data_science_mcp.kernels._runner",
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 data_science_mcp/kernels/kernel_verifier.py
low Dynamic code execution in packaging/dev tooling (tests/test_launch.py)MTC-SRC-001
In a packaging/dev/install script (shipped, but not the server runtime) (`tests/test_launch.py:2`): Evaluating strings as code is the most direct RCE primitive; if any tool input reaches it, the server executes attacker-chosen code. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.
Evidence: ed launcher + benchmark eval (CONCEPT:DS-AHE.trainer.concept-4/006). Config builders and the ``accelerate launch`` argv
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 tests/test_launch.py
low Shell/command execution in packaging/dev tooling (scripts/security_sanitizer.py)MTC-SRC-002
In a packaging/dev/install script (shipped, but not the server runtime) (`scripts/security_sanitizer.py:136`): 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: try: result = subprocess.run( ["git", "ls-files", "--cached", "--others", "--exclude-standard"],
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 scripts/security_sanitizer.py
low Shell/command execution in packaging/dev tooling (tests/conftest.py)MTC-SRC-002
In a packaging/dev/install script (shipped, but not the server runtime) (`tests/conftest.py:81`): 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: ngine.sock") proc = subprocess.Popen( [binary, "--socket-path", sock], stdout=subprocess.PIPE,
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 tests/conftest.py
Tools 38
Each tool and what it can reach — statically extracted from the published source.
build_training_datasetno sensitive capabilitycompose_rewardno sensitive capabilitycross_validateno sensitive capabilitycurate_corpusno sensitive capabilitydataset_lineageno sensitive capabilitydecontaminate_corpusno sensitive capabilitydedup_corpusno sensitive capabilitydeep_train_predictno sensitive capabilitydescribe_datasetno sensitive capabilityds_specialize_kernelno sensitive capability
Show 28 more tools ↓
evaluate_modelno sensitive capabilityevolve_model_classno sensitive capabilityfit_modelno sensitive capabilitygenerate_interpretability_testsno sensitive capabilityget_pareto_frontierno sensitive capabilitygrade_responseno sensitive capabilityload_datasetno sensitive capabilitymerge_adapters_tiesno sensitive capabilitypredictno sensitive capabilityprepare_pretrain_datano sensitive capabilitypretrain_modelno sensitive capabilityquant_derivativesno sensitive capabilityquant_forensicno sensitive capabilityquant_market_makingno sensitive capabilityquant_microstructureno sensitive capabilityquant_signalsno sensitive capabilityquant_sizingno sensitive capabilityquant_statespaceno sensitive capabilityquant_validationno sensitive capabilityrank_modelsno sensitive capabilityrun_interpretability_suiteno sensitive capabilitysplit_datasetno sensitive capabilitytrain_dpono sensitive capabilitytrain_grpono sensitive capabilitytrain_ppono sensitive capabilitytrain_rewardno sensitive capabilitytrain_sftno sensitive capabilitytrain_tokenizerno sensitive capability
What this scan could not see
- Tools were statically extracted from the published source (38 recovered), not enumerated from a running server. Tool-poisoning, Unicode-smuggling, capability and toxic-flow analysis ran on this inferred surface, but a mis-parsed registration could be missed or mis-attributed, so tool-derived findings are capped below “confirmed”. To grade the real runtime surface, scan the running server: --command "npx -y <package>".
Versions 1
Scan history per published version. The engine is deterministic — the same version always yields the same score, so a changed score means the package itself changed.
| Version | Score | Findings | Engine | Scanned |
|---|---|---|---|---|
v1.2.0 latest |
A 92/100 | 8 | 1.13.0 | 2026-08-25 |
Embed this score
Show this server's live Trust Score in your README, docs or website. The badge is served straight from the registry and updates automatically after every rescan — no API key needed. It links back to this page, so anyone who sees the grade can also read the findings behind it instead of taking a number on faith.
Verify this score yourself
The score above is reproducible: the same package version always yields the same result. Run it locally or over the free API — no account, no LLM, fully deterministic.
npx mcptrustchecker scan data-science-mcp --online --registry pypiMore in Developer Tools
Hotel booking MCP server — 300K+ properties, real confirmation numbers, loyalty programs. Builders monetize every booking via Stripe Connect. The first MCP server that completes real hotel reservations inside AI conversations.
Manage AdGuard Home through AI assistants
Read-only Azure DevOps for MCP clients using only your existing browser session — no PAT, no Azure CLI. Browse work items, pull requests, comments, attachments and Artifacts feeds across every project, repo and feed you can access.
MCP server for Adobe Experience Manager Assets integration development
Servidor MCP para el tiempo oficial de España (API pública OpenData de AEMET). Predicción, observación y avisos como herramientas MCP tipadas.
A standalone MCP stdio bridge for Affinity by Canva's local MCP SSE server.