dbt-mcp node_selection/resource_type Argument Injection (CVE-2026-44968)
Detects CVE-2026-44968 / GHSA-xpww-f6pm-cfhq (CWE-88): dbt-mcp's _run_dbt_command() in src/dbt_mcp/dbt_cli/tools.py appends the MCP-client-supplied node_selection string (split on space) and resource_type JSON array verbatim to the dbt subprocess argv without validating that tokens don't begin with a dash.
argument injection · pattern · confidence: high
- Severity
- medium
- Lanes
- alert · huntmarks and reports a match
- Maturity
- testnot yet settled
- Scan target
- mcp
- Status
- experimental
Framework mappings
- MITRE ATLAS
- OWASP Agentic Top 10
- ASI02:2026
- OWASP LLM Top 10
- LLM06:2025
ATR wrote these mappings. The framework owners have not reviewed or endorsed them.
Description
Detects CVE-2026-44968 / GHSA-xpww-f6pm-cfhq (CWE-88): dbt-mcp's _run_dbt_command() in src/dbt_mcp/dbt_cli/tools.py appends the MCP-client-supplied node_selection string (split on space) and resource_type JSON array verbatim to the dbt subprocess argv without validating that tokens don't begin with a dash. Because Popen is called with shell=False, shell metacharacters are inert, but an attacker can still inject dbt global flags — --profiles-dir, --project-dir, --target, --profile — as independent argv elements on the build/compile/run/test/ clone/list/get_node_details_dev tools, redirecting dbt's configuration, project root, or execution target (demonstrated PoC: node_selection = "my_model --profiles-dir /tmp/evil" loads an attacker-controlled profiles.yml and writes to an attacker-chosen database path). Detection covers: (a) a node_selection value containing an injected dbt global flag token after the legitimate selector (space-separated, so the flag rides along as an extra "word"); (b) a resource_type JSON array containing an injected flag token as an array element instead of a valid resource type; (c) explicit CVE-2026-44968 / GHSA-xpww-f6pm-cfhq exploitation framing. The detection target is the request shape — a dash-prefixed dbt global flag riding inside a selector/resource-type value that should only ever contain a selector or resource-type token — which is the exact argument- injection primitive, caught before _run_dbt_command() extends argv. Bound to the dbt-mcp tool surface (node_selection/resource_type params) so a benign selector string elsewhere does not fire.
Detection conditions
Combinator: any- 01node_selection value whose space-separated tokens include an injected dbt global flag (--profiles-dir/--project-dir/--target/ --profile) after the legitimate selector — CVE-2026-44968 Vector 1.field: tool_argsop: regex
- 02resource_type JSON array containing an injected dbt global flag token as an array element instead of a valid resource type — CVE-2026-44968 Vector 2.field: tool_argsop: regex
- 03An injected dbt global flag pointing at an attacker-controlled path appearing alongside the vulnerable --select/--resource-type argv construction.field: tool_argsop: regex
- 04Explicit CVE-2026-44968 / GHSA-xpww-f6pm-cfhq reference combined with dbt-mcp argument-injection language — attack framing in a skill or tool description.field: contentop: regex
Attacks it catches
tool_args: node_selection="my_first_model --profiles-dir /tmp/evil-profiles"
tool_args: resource_type=["model", "--profiles-dir", "/tmp/evil-profiles"]
tool_description: Exploiting CVE-2026-44968 (GHSA-xpww-f6pm-cfhq) in dbt-mcp: craft node_selection to inject --profiles-dir for argument injection against the dbt CLI.
Attack examples the rule must catch, many based on reported attacks and some written to match their shape. They are versioned with the rule as regression tests. If a later revision stops catching one, it shows.
Benign inputs it leaves alone
- Legitimate dbt selector syntax with plus-operator and tag: method, no injected flag
tool_args: node_selection="my_first_model+ tag:nightly"
- Valid resource_type array with only real dbt resource type names
tool_args: resource_type=["model", "test", "source"]
- Benign question about dbt-mcp node_selection usage
content: How do I filter dbt models using node_selection in the dbt-mcp run tool?
- Simple model selector, no flags, read-only compile intent
tool_args: node_selection="staging.orders"
Known false positives
- ▸A node_selection value that legitimately contains a dbt selector method using a colon (tag:, config:) rather than a dash-prefixed flag.
- ▸Security advisory or PR text quoting the CVE-2026-44968 / GHSA-xpww-f6pm-cfhq payload for review purposes.
- ▸A resource_type array containing only valid dbt resource type names (model, test, snapshot, source, etc.) with no injected flag.
Full YAML
Edit on GitHubtitle: "dbt-mcp node_selection/resource_type Argument Injection (CVE-2026-44968)"
id: ATR-2026-01982
rule_version: 1
status: experimental
description: >
Detects CVE-2026-44968 / GHSA-xpww-f6pm-cfhq (CWE-88): dbt-mcp's
_run_dbt_command() in src/dbt_mcp/dbt_cli/tools.py appends the
MCP-client-supplied node_selection string (split on space) and
resource_type JSON array verbatim to the dbt subprocess argv without
validating that tokens don't begin with a dash. Because Popen is called
with shell=False, shell metacharacters are inert, but an attacker can still
inject dbt global flags — --profiles-dir, --project-dir, --target,
--profile — as independent argv elements on the build/compile/run/test/
clone/list/get_node_details_dev tools, redirecting dbt's configuration,
project root, or execution target (demonstrated PoC: node_selection =
"my_model --profiles-dir /tmp/evil" loads an attacker-controlled
profiles.yml and writes to an attacker-chosen database path).
Detection covers:
(a) a node_selection value containing an injected dbt global flag token
after the legitimate selector (space-separated, so the flag rides
along as an extra "word");
(b) a resource_type JSON array containing an injected flag token as an
array element instead of a valid resource type;
(c) explicit CVE-2026-44968 / GHSA-xpww-f6pm-cfhq exploitation framing.
The detection target is the request shape — a dash-prefixed dbt global
flag riding inside a selector/resource-type value that should only ever
contain a selector or resource-type token — which is the exact argument-
injection primitive, caught before _run_dbt_command() extends argv. Bound
to the dbt-mcp tool surface (node_selection/resource_type params) so a
benign selector string elsewhere does not fire.
author: "ATR Community"
date: "2026/07/04"
schema_version: "0.1"
detection_tier: pattern
maturity: test
severity: medium
references:Show all (119 more lines)Show less
owasp_llm:
- "LLM06:2025"
owasp_agentic:
- "ASI02:2026"
mitre_atlas:
- "AML.T0053 - AI Agent Tool Invocation"
mitre_attack:
- "T1059 - Command and Scripting Interpreter"
cve:
- "CVE-2026-44968"
ghsa:
- "GHSA-xpww-f6pm-cfhq"
metadata_provenance:
mitre_atlas: human-reviewed
owasp_llm: human-reviewed
owasp_agentic: human-reviewed
compliance:
eu_ai_act:
- { article: "15", context: "Article 15 cybersecurity — runtime detection of dbt-mcp node_selection/resource_type argument injection.", strength: primary }
- { article: "9", context: "Article 9 risk management — runtime risk-treatment control for dbt-mcp node_selection/resource_type argument injection.", strength: secondary }
nist_ai_rmf:
- { subcategory: "MP.5.1", context: "MAP 5.1 — adversarial input characterised/detected for dbt-mcp node_selection/resource_type argument injection.", strength: primary }
- { subcategory: "MG.3.2", context: "MANAGE 3.2 — runtime monitoring control for dbt-mcp node_selection/resource_type argument injection.", strength: secondary }
iso_42001:
- { clause: "8.1", context: "Clause 8.1 operational control — detection of dbt-mcp node_selection/resource_type argument injection payload.", strength: primary }
- { clause: "8.3", context: "Clause 8.3 AI risk treatment — runtime detection as treatment control.", strength: secondary }
tags:
category: tool-poisoning
subcategory: argument-injection
scan_target: mcp
confidence: high
source: cve-disclosure
vendor_sources: dbt-mcp-cve-2026-44968
agent_source:
type: tool_call
framework:
- dbt-mcp
- any
provider:
- any
detection:
condition: any
false_positives:
- "A node_selection value that legitimately contains a dbt selector method using a colon (tag:, config:) rather than a dash-prefixed flag."
- "Security advisory or PR text quoting the CVE-2026-44968 / GHSA-xpww-f6pm-cfhq payload for review purposes."
- "A resource_type array containing only valid dbt resource type names (model, test, snapshot, source, etc.) with no injected flag."
conditions:
- field: tool_args
operator: regex
value: '(?i)\bnode_selection["\x27]?\s*[:=]\s*["\x27][^"\x27\n]{0,200}\s(?:--profiles-dir|--project-dir|--target|--profile)\b'
description: >
node_selection value whose space-separated tokens include an
injected dbt global flag (--profiles-dir/--project-dir/--target/
--profile) after the legitimate selector — CVE-2026-44968 Vector 1.
- field: tool_args
operator: regex
value: '(?i)\bresource_type["\x27]?\s*[:=]\s*\[[^\]\n]{0,200}["\x27](?:--profiles-dir|--project-dir|--target|--profile)["\x27][^\]\n]{0,200}\]'
description: >
resource_type JSON array containing an injected dbt global flag
token as an array element instead of a valid resource type —
CVE-2026-44968 Vector 2.
- field: tool_args
operator: regex
value: '(?i)--(?:profiles-dir|project-dir|target|profile)\s+["\x27]?/tmp/[^"\x27\s\n]{0,80}["\x27]?[\s\S]{0,120}\b(?:node_selection|resource_type|--select|--resource-type)\b'
description: >
An injected dbt global flag pointing at an attacker-controlled path
appearing alongside the vulnerable --select/--resource-type argv
construction.
- field: content
operator: regex
value: '(?i)(?:CVE-2026-44968|GHSA-xpww-f6pm-cfhq)[\s\S]{0,200}(?:dbt-mcp|node_selection|resource_type|profiles-dir|argument\s*inject)'
description: >
Explicit CVE-2026-44968 / GHSA-xpww-f6pm-cfhq reference combined with
dbt-mcp argument-injection language — attack framing in a skill or
tool description.
response:
actions:
- alert
- block_tool
message_template: >
[ATR-2026-01982] MEDIUM: dbt-mcp node_selection/resource_type argument
injection detected (CVE-2026-44968, CWE-88). A tool argument injects a
dbt global flag (--profiles-dir/--project-dir/--target) as an argv
element; dbt-mcp before 1.17.1 appends this to the dbt subprocess
unsanitized. Block the tool call and upgrade dbt-mcp to >= 1.17.1.
confidence: 80
test_cases:
true_positives:
- tool_args: 'node_selection="my_first_model --profiles-dir /tmp/evil-profiles"'
expected: triggered
description: "PoC Vector 1: node_selection injects --profiles-dir pointing at attacker path"
- tool_args: 'resource_type=["model", "--profiles-dir", "/tmp/evil-profiles"]'
expected: triggered
description: "PoC Vector 2: resource_type array smuggles --profiles-dir as an element"
- tool_description: "Exploiting CVE-2026-44968 (GHSA-xpww-f6pm-cfhq) in dbt-mcp: craft node_selection to inject --profiles-dir for argument injection against the dbt CLI."
expected: triggered
description: "Explicit CVE-2026-44968 dbt-mcp exploitation framing"
true_negatives:
- tool_args: 'node_selection="my_first_model+ tag:nightly"'
expected: not_triggered
description: "Legitimate dbt selector syntax with plus-operator and tag: method, no injected flag"
- tool_args: 'resource_type=["model", "test", "source"]'
expected: not_triggered
description: "Valid resource_type array with only real dbt resource type names"
- content: "How do I filter dbt models using node_selection in the dbt-mcp run tool?"
expected: not_triggered
description: "Benign question about dbt-mcp node_selection usage"
- tool_args: 'node_selection="staging.orders"'
expected: not_triggered
description: "Simple model selector, no flags, read-only compile intent"