mcp-vulnerability-dataset / BEGINNER_GUIDE.md
XuViewer's picture
Use the full split and document the frozen data cutoff
f2ff158 verified
|
Raw History Blame Contribute Delete
22.3 kB

A beginner's guide to reading the MCP Vulnerability Dataset

This guide explains all 18 top-level fields using the first record, CVE-2025-10193, a DNS rebinding vulnerability in the Neo4j Cypher MCP server. No prior knowledge of MCP or vulnerability research is assumed.

Open the native Dataset Viewer alongside this guide. For compact definitions, see the technical field reference; for classification boundaries, see TAXONOMY.md.

Start with the story

MCP (Model Context Protocol) provides a common way for applications, including AI assistants, to connect to tools and data sources. An MCP client consumes these capabilities; an MCP server provides them. A “server” is a software role: it may run on your own computer, rather than on a remote machine.

In this example, the Neo4j Cypher MCP server exposes tools for working with a Neo4j database. The advisory describes a malicious website reaching a vulnerable MCP service running locally and triggering unauthorized tool invocations. The attack requires the user to visit the malicious website and remain there long enough for DNS rebinding to work.

DNS translates a domain name into an IP address. In a DNS rebinding attack, an attacker changes the address associated with a name so that a web page can reach a different destination, such as a local service. This exploits a mismatch between the browser's view of a web origin and the service actually reached.

A useful reading of this record is:

Malicious web content or a DNS response reaches across the boundary between a web origin and a local service, enabling unauthorized Neo4j tool calls.

This does not establish that every Neo4j installation is affected, or that the attack necessarily executes arbitrary operating-system commands. The affected product, version range and disclosed prerequisites matter. The project security advisory is the primary reference for this example.

1–7: Identify the vulnerability and its scope

Field Example value How to read it
1. cve_id CVE-2025-10193 The public identifier used to find this vulnerability across sources. It is not a severity score.
2. title Neo4j Cypher MCP server is vulnerable to DNS rebinding attacks A short upstream headline. Use it to scan records quickly.
3. description The upstream account of unauthorized local tool calls through DNS rebinding Read for the reported behavior, affected system and attack prerequisites. It is distinct from our mechanism interpretation.
4. date_published 2025-09-11T14:05:30.592Z The official first-publication timestamp. Z means UTC. This is not the discovery date, the date the bug was introduced, or our collection date.
5. affected Product, package and version objects Use this to identify which software and versions the upstream record lists as affected.
6. upstream_weaknesses CWE-346, Origin Validation Error Weakness categories supplied by upstream sources. They are not automatically generated from our taxonomy.
7. upstream_severity_metrics CVSS 4.0, score 7.4, severity HIGH Upstream severity assessments and their assumptions. These are not scores assigned by this dataset's reviewers.

CVE identifies an individual vulnerability; CWE identifies a kind of weakness. Many CVEs can share a CWE. Here, origin validation means checking where a request comes from and whether that origin should be trusted. “Upstream” means information preserved from the disclosure sources, rather than a new project judgment.

Reading affected

This field is a list because a record can cover multiple products or packages. For the example, its populated fields are:

Nested field Value Meaning
vendor neo4j Vendor or maintaining entity named by the source
product neo4j-cypher MCP server The affected product
packageName mcp-neo4j-cypher The package name to match against software inventories
collectionURL https://pypi.python.org The package distribution platform
repo The project's GitHub URL The source repository
versions[].version 0.2.2 Lower bound of this version range
versions[].lessThanOrEqual 0.3.1 Inclusive upper bound
versions[].status affected Status assigned to that range
versions[].versionType semver The version-comparison scheme
defaultStatus unaffected The source's default for versions outside its explicit entries; not our independent verification of every other version

Read this range as 0.2.2 through 0.3.1, inclusive, not just its two endpoints. The advisory separately identifies 0.4.0 as a patched version. Do not infer a fixed version merely by incrementing an affected version number.

In the Viewer, you may also see optional keys such as modules or versions[].lessThan with null values. Standard Parquet structs share one schema across rows, so absent keys are padded with null. JSON/JSONL downloads retain original key presence. A padded null does not add a new upstream claim.

Reading upstream_weaknesses

Within each descriptions object, cweId is the identifier, description is its name or description, lang identifies the language, and type indicates the classification system. In this example, lang: en and type: CWE are metadata, not additional weaknesses.

Reading upstream_severity_metrics

New readers can start with the score and the conditions that produced it:

Metric Example value Plain-language meaning
baseScore / baseSeverity 7.4 / HIGH A standardized high-severity assessment, not a 74% success probability
attackVector NETWORK Exploitation involves network access; this does not mean the local service must be directly exposed to the public internet
attackComplexity HIGH There are additional technical difficulties in exploitation
attackRequirements PRESENT Specific deployment or execution conditions must hold
privilegesRequired NONE The assessment does not require the attacker to already hold target-system privileges
userInteraction ACTIVE The user must participate; read the description for this case's concrete requirements
vulnConfidentialityImpact HIGH Potentially substantial exposure of information that should be protected
vulnIntegrityImpact HIGH Potentially substantial unauthorized changes to information or system state
vulnAvailabilityImpact NONE The assessment does not assign an availability impact, such as making the service unusable
subConfidentialityImpact, subIntegrityImpact, subAvailabilityImpact NONE No impact is assigned to subsequent systems in this assessment
version / vectorString 4.0 / CVSS:4.0/AV:N/... The scoring-standard version and compact encoding of the assessment

Supplemental fields add context: Automatable concerns automation of the attack; Recovery concerns recovery after exploitation; Safety concerns human safety; valueDensity concerns the concentration of resources affected; vulnerabilityResponseEffort concerns response effort; and providerUrgency records the provider's urgency assessment. NOT_DEFINED means unspecified, not “no risk.” The outer format names the scoring format, while scenarios describes the assessment's scenario and language (GENERAL and en here). Unused CVSS-version objects may appear as null in Parquet.

Some other records retain NVD-format metric groups. For those records, the list contains one object with keys such as cvssMetricV31 or ssvcV203. Read a CVSS assessment at upstream_severity_metrics[0]["cvssMetricV31"][i]["cvssData"]; its surrounding source and type identify the upstream assessor where supplied. SSVC is a separate prioritization assessment, not a CVSS score. No metric groups, provider names or scores are discarded to force them into this example's shape.

CVSS describes severity, not the actual risk of every installation. Deployment, permissions and accessible data still matter. Consult the official CVSS 4.0 specification for exact metric definitions.

8–11: Understand the project's classification

Field Example value Question it answers
8. vulnerability_type DNS Rebinding What is the primary evidenced failure mechanism?
9. component server In which component role does the defect occur? Here, the role provides MCP tools.
10. root_cause ["Protocol/Interaction Boundary Controls"] Which kind of security control failed? Here, controls at a protocol or interaction boundary.
11. classification_confidence high for all three classifications How strongly does the available evidence support each classification?

These labels describe different dimensions. The type is DNS rebinding, the component is an MCP server, and the root cause concerns insufficient controls between web-origin access and the local service.

root_cause is a list because more than one supported control failure can apply. It contains positive findings only: an omitted cause has not necessarily been ruled out. For classification boundaries, consult TAXONOMY.md.

Inside classification_confidence, component_bucket refers to component; vulnerability_type and root_cause refer to their corresponding labels. Confidence is qualitative, not a calibrated probability or proof of an exploit. High severity means potentially serious consequences; high classification confidence means stronger support for the assigned label. Do not confuse them.

12–16: Trace the mechanism and its relationship to MCP

Field Example value How to read it
12. mechanism_relation traditional_weakness_at_mcp_integration_boundary A familiar software-security weakness appears where MCP connects to another system. Its presence in an MCP product does not make it unique to MCP.
13. model_necessity not_required_for_documented_primitive The documented attack mechanism does not require a large language model to interpret content, make a decision or call a tool.
14. sensitive_operation Local Neo4j tool invocation The security-sensitive capability reached by the attack. Do not silently expand “tool invocation” into arbitrary operating-system command execution.
15. trust_boundary Web origin → local service The separation between differently trusted actors or capabilities that the attack crosses.
16. untrusted_entry Malicious web page or DNS response The attacker-controlled or insufficiently trusted input entering the system. Here, it is not a chat prompt.

Read the last three fields as a short causal path:

Malicious web page / DNS response → web-to-local-service boundary → local tool invocation

Crossing a trust boundary is normal in many legitimate operations. The security failure is a missing, bypassed or ineffective check at that crossing. An external website should not automatically receive permission to operate local tools.

model_necessity does not ask whether the product uses an AI model. It asks whether the documented exploit mechanism depends on one. It also does not rule out an alternative model-assisted attack path. This distinction helps separate traditional software defects in AI products from attacks that depend on model behavior, such as a documented prompt-injection chain.

17. evidence: Find the supporting sources

This record has two evidence objects:

ID Source Role in this record
E0 CVE record Basic upstream information such as description, affected versions and severity
E1 Project security advisory Evidence used for mechanism interpretation and classification

Every evidence object uses these fields:

Nested field Meaning
id A reference ID unique within this record; another CVE may also have its own E1
kind Source type, such as cve_record, security_advisory, commit, source_code or issue_report
url Primary source link
related_urls Additional links recorded for the source artifact; an empty list means none are listed
supports Record fields that this evidence contributes to supporting
summary A short project description of the source's purpose, not a reproduced article

Supporting sources form a collective basis. A source listed for a field does not necessarily prove the entire attack chain on its own. Links may point to upstream content that changes after the saved dataset snapshot. Original source snapshots and hashes remain in the local audit archive.

18. reasoning: Understand how evidence became a judgment

Evidence answers “What did you consult?” Reasoning answers “Why does it support this assessment?” The object is keyed by the field being assessed.

For example, the model_necessity rationale says that DNS rebinding operates between the web page and the local service without requiring model decisions. This is a mechanism inference based on the advisory, not a separate official statement that every possible attack is model-independent.

Nested field Meaning
basis Whether the assessment is a source-supported observation, an interpretation or an unresolved judgment
reason The explanation for the assigned value
limitation What the assessment does not establish
evidence_ids IDs to look up in this same record's evidence list

Important basis values are:

  • source_supported_at_stated_granularity: the saved source supports an observation at the stated level of detail.
  • analyst_mechanistic_inference: the project interprets the mechanism from available evidence.
  • analyst_classification: the project applies its taxonomy to the evidence.
  • insufficient_or_conflicting_evidence: available information does not support a definite answer; this value occurs in other records.

The basis marker is essential when citing the dataset: an official disclosure, a project classification and a mechanistic inference are not interchangeable. The annotations were developed through AI-assisted static review, without independent human validation. Inclusion does not claim successful runtime reproduction. This example's rationale is shared across several mechanism fields; it should be read with each field's scope and supporting evidence, not as a separate experiment for each field.

A practical reading workflow

  1. Read title and description to identify the behavior and prerequisites.
  2. Inspect affected to match the product, package and version range.
  3. Read vulnerability_type, component and root_cause as separate dimensions.
  4. Connect untrusted_entry, trust_boundary and sensitive_operation into an attack path.
  5. Check model_necessity and mechanism_relation before making claims about AI-specific or MCP-specific risk.
  6. Follow reasoning to evidence for any conclusion you want to rely on or cite.

After reading this example, a reasonable explanation is:

Certain versions of the Neo4j MCP tool service allow a malicious website to use DNS rebinding to reach local tools without authorization. The documented mechanism does not depend on a language model. The classification is supported by disclosure evidence, but the dataset does not claim independent exploitation.

Working with the records

For a quick local inspection of the downloaded JSONL, Python's standard library is sufficient:

import json

with open("records.jsonl", encoding="utf-8") as source:
    records = [json.loads(line) for line in source if line.strip()]

record = records[0]
print(record["cve_id"], record["vulnerability_type"])

# Follow a judgment to its sources.
evidence_by_id = {item["id"]: item for item in record["evidence"]}
judgment = record["reasoning"]["model_necessity"]
print(judgment["basis"], judgment["reason"])
for evidence_id in judgment["evidence_ids"]:
    print(evidence_by_id[evidence_id]["url"])

# Keep unknowns visible rather than silently counting them as negative findings.
unknown_boundaries = [r["cve_id"] for r in records if r["trust_boundary"] is None]
print("Unestablished trust boundaries:", unknown_boundaries)

null means unestablished, unspecified or, for padded optional Parquet keys, absent in the source object. It does not mean false or safe. An empty upstream list means no values were available in the exported snapshot. Check the field scope and reasoning to distinguish these cases.

Use CVE IDs as record identifiers. A CVE record count is not necessarily a count of independent root causes, affected installations or successful attacks. A multi-label root-cause count can exceed the number of records. These 791 records form a frozen research cohort, not an exhaustive census or a representative random sample of the MCP ecosystem. The full split contains the whole cohort.

For reproducible research, record the dataset revision and consult the taxonomy version in the README. Preserve source attribution and the license notices when redistributing data. For deployment decisions, recheck current upstream advisories rather than treating a frozen dataset snapshot as a live security feed.

Technical field reference

All rows share this schema; omitted audit fields are documented globally in README.

Field Meaning
cve_id Stable CVE identifier and join key
title, description Upstream title and English description; not our mechanism conclusion
date_published Official first-publication timestamp, not NVD ingestion date
affected Upstream product/vendor/version objects, preserved without flattening version ranges
upstream_weaknesses Upstream weakness objects; not automatically mapped from our taxonomy
upstream_severity_metrics Upstream severity/assessment objects (including CVSS and, where present, SSVC); absent metrics remain an empty list
vulnerability_type Primary mechanism-first taxonomy label
component Role of affected component; null if unresolved
root_cause Positive supported root labels; null when none established, not absence of a flaw
classification_confidence Original per-field qualitative confidence; component_bucket refers to component
untrusted_entry Attacker-controlled or insufficiently trusted entry
sensitive_operation Security-sensitive operation reached
trust_boundary Trust or privilege boundary; null when unestablished
model_necessity Model necessity for the documented primitive/chain
mechanism_relation Traditional reuse, integration boundary, agent composition, or null
evidence Source objects described below
reasoning Field-keyed basis, reason, limitations and evidence_ids

Evidence objects

id is unique within a record. kind distinguishes cve_record, security_advisory, commit, source_code, issue_report and reference. url is the primary source; related_urls retains additional links recorded for the same local evidence artifact. supports lists record fields. summary explains the source's purpose. reasoning.<field>.evidence_ids joins to these objects. Background references can have an empty supports list. No raw evidence bodies or local filesystem paths are exported. Evidence locators may be added when verified; this release does not invent line numbers or function names for sources without reliable locators.

Unknown and basis semantics

JSON null means unknown/unestablished, never false. Empty upstream lists mean no values were available in this exported source snapshot. Root arrays list only supported positives and do not exhaust all possible causes. Unknown mechanisms retain a reason and their evaluated evidence rather than guessing a category.

source_supported_at_stated_granularity: supported by saved source at that scope. analyst_mechanistic_inference: our interpretation of the mechanism. insufficient_or_conflicting_evidence: insufficient/conflicting source evidence. analyst_classification: project taxonomy judgment, not an official upstream label.

Model values and mechanism relation values are preserved verbatim from v1. not_required_for_documented_primitive means no model is needed for the described primitive; it does not rule out model-assisted exploitation.

MCP relevance, taxonomy version, individual review date, reviewer identity and runtime-reproduction flags are intentionally absent from each public row. Complete values remain in the local audit; README supplies method and scope.

Parquet representation

The Viewer uses standard Parquet list/struct types, without JSON extension types. Absent optional object keys are represented as null in Parquet so that all rows share a schema. The JSON/JSONL downloads preserve the original key presence. Existing values, explicit nulls, evidence bindings and record order are unchanged.

Upstream metric shapes

upstream_severity_metrics is always a list. CVE-format metric objects retain keys such as cvssV3_1, cvssV4_0, format and scenarios. In 42 records, the frozen catalog instead stores an NVD-format dictionary of metric groups. That entire dictionary is wrapped in a one-element list, preserving keys such as cvssMetricV31 and ssvcV203, all nested values and provider attribution. This structural normalization is reversible; it is not a reassessment or score mapping. The local audit retains both forms. Missing source attribution is not inferred.

The full cohort may add optional nested keys relative to the ten-record preview; the same 18 top-level fields and their meanings remain unchanged.