Cypher/Graph-Query Injection via Unsanitized node_labels or group_ids Field

Detects an agent-facing graph-memory or knowledge-graph tool call where a node_labels or group_ids parameter contains a graph-query metacharacter (a closing backtick, parenthesis, or brace) immediately followed by a Cypher/graph-query clause keyword (MATCH, MERGE, CREATE, CALL db., or DETACH DELETE). Legitimate node_labels/group_ids values are short identifier-style strings (e.g. "Person", "team-alpha") with no query syntax; an escape character followed by a full clause is the graph-database equivalent of a classic SQL injection payload, generalized to the Cypher query language used by agent temporal/knowledge-graph memory backends. Mined from GHSA-gg5m-55jj-8m5g (Graphiti PR #1312, CVE-2026-32247): "the vulnerability existed in how node_labels and group_ids parameters were handled before being used to construct Cypher, Lucene, or RedisSearch query fragments. Without proper validation, attackers could inject malicious Cypher syntax through these parameters," patched by "rejecting unsafe node_labels and group_ids before search code execution" and "extending protection across write-path label interpolation." Deliberately distinct from ATR-2026-00570 (generic SQL injection), which targets relational SQL keywords (OR/UNION/SELECT) rather than the Cypher-specific DETACH DELETE / MATCH / CALL db. vocabulary this rule targets.

建議回應

參考資料

1. 01

   node_labels/group_ids value breaks out of its identifier string with a graph-query metacharacter followed by a Cypher clause keyword

   欄位: content運算子: regex

攻擊範例（規則會觸發）

1. {"node_labels": ["Person`) DETACH DELETE (n) RETURN 1 //"]}

2. {"group_ids": ["grp1`}) MATCH (a)-[r]-(b) DELETE r //"]}

3. node_labels: ["Foo) CALL db.dropDatabase() //"]

以上為真實攻擊 payload 的脫敏版本，與規則一同版本化，作為 regression test——確保未來的修訂不會悄悄漏掉它們。

正常樣本（規則不會觸發）

1. Ordinary plain-identifier node_labels values

   {"node_labels": ["Person", "Employee"]}

2. Ordinary plain-identifier group_ids values

   {"group_ids": ["team-alpha", "team-beta"]}

3. Legitimate display name containing a parenthesis but no Cypher clause keyword following it

   {"node_labels": ["Company (Subsidiary)"]}

已知誤報情境

• ▸Ordinary node_labels or group_ids values that are plain identifier strings (Person, Employee, team-alpha) with no escape characters or query keywords
• ▸A node_labels/group_ids value that happens to contain a parenthesis as part of a legitimate display name, without a following Cypher clause keyword

title: "Cypher/Graph-Query Injection via Unsanitized node_labels or group_ids Field"
id: ATR-2026-02143
rule_version: 1
status: experimental
description: >
  Detects an agent-facing graph-memory or knowledge-graph tool call where a
  node_labels or group_ids parameter contains a graph-query metacharacter
  (a closing backtick, parenthesis, or brace) immediately followed by a
  Cypher/graph-query clause keyword (MATCH, MERGE, CREATE, CALL db., or
  DETACH DELETE). Legitimate node_labels/group_ids values are short
  identifier-style strings (e.g. "Person", "team-alpha") with no query
  syntax; an escape character followed by a full clause is the graph-database
  equivalent of a classic SQL injection payload, generalized to the Cypher
  query language used by agent temporal/knowledge-graph memory backends.
  Mined from GHSA-gg5m-55jj-8m5g (Graphiti PR #1312, CVE-2026-32247):
  "the vulnerability existed in how node_labels and group_ids parameters were
  handled before being used to construct Cypher, Lucene, or RedisSearch query
  fragments. Without proper validation, attackers could inject malicious
  Cypher syntax through these parameters," patched by "rejecting unsafe
  node_labels and group_ids before search code execution" and "extending
  protection across write-path label interpolation." Deliberately distinct
  from ATR-2026-00570 (generic SQL injection), which targets relational
  SQL keywords (OR/UNION/SELECT) rather than the Cypher-specific
  DETACH DELETE / MATCH / CALL db. vocabulary this rule targets.
author: "ATR Community (CVE sweep)"
date: "2026/07/11"
schema_version: "0.1"
detection_tier: pattern
maturity: experimental
severity: high

references:
  owasp_llm:
    - "LLM01:2025 - Prompt Injection"
  owasp_agentic:
    - "ASI02:2026 - Tool Misuse and Exploitation"
  mitre_attack:
    - "T1565.001 - Data Manipulation: Stored Data Manipulation"
  mitre_atlas:
    - "AML.T0053 - LLM Plugin Compromise"
  cve:
    - "CVE-2026-32247"
  external:
    - "https://github.com/getzep/graphiti/pull/1312"
    - "https://github.com/getzep/graphiti/releases/tag/v0.28.2"

metadata_provenance:
  owasp_llm: human-reviewed
  owasp_agentic: human-reviewed
  mitre_attack: human-reviewed
  mitre_atlas: human-reviewed
  cve: nvd-sync

compliance:
  eu_ai_act:
    - article: "15"
      context: "Article 15 (accuracy, robustness and cybersecurity) requires high-risk AI systems to resist unauthorised attempts to alter their behaviour or environment; this rule detects a graph-query injection through an agent memory tool's node_labels/group_ids parameter."
      strength: primary
    - article: "10"
      context: "Article 10 (data and data governance) requires control over the data an AI system processes; this rule detects tampering with the integrity of the agent's knowledge-graph memory store via query injection."
      strength: secondary
  nist_ai_rmf:
    - subcategory: "MG.2.3"
      context: "Treating graph-query injection as an identified AI risk requires active runtime countermeasures; this detection rule is the primary risk treatment implementation."
      strength: primary
    - subcategory: "MP.5.1"
      context: "Identifying node_labels/group_ids graph-query injection as an AI risk to be catalogued in the organizational risk register."
      strength: secondary
  iso_42001:
    - clause: "8.1"
      context: "ISO/IEC 42001 Clause 8.1 (operational planning and control) is operationalised by this rule's detection of graph-query injection attempts."
      strength: primary
    - clause: "8.3"
      context: "ISO/IEC 42001 Clause 8.3 (AI risk treatment) is implemented via this rule's runtime detection of the injection attempt."
      strength: secondary

tags:
  category: data-poisoning
  subcategory: graph-query-injection
  scan_target: mcp
  confidence: medium

agent_source:
  type: llm_io
  framework:
    - any
  provider:
    - any

detection:
  condition: any
  conditions:
    - field: content
      operator: regex
      value: '(?i)"?(?:node_labels|group_ids)"?\s*[:=]\s*\[?\s*["\x27][^"\x27]{0,40}[`)\}][\s\S]{0,30}\b(?:detach\s+delete|match\s*\(|merge\s*\(|call\s+db\.|create\s*\()'
      description: "node_labels/group_ids value breaks out of its identifier string with a graph-query metacharacter followed by a Cypher clause keyword"
  false_positives:
    - "Ordinary node_labels or group_ids values that are plain identifier strings (Person, Employee, team-alpha) with no escape characters or query keywords"
    - "A node_labels/group_ids value that happens to contain a parenthesis as part of a legitimate display name, without a following Cypher clause keyword"

response:
  actions:
    - block_input
    - alert
  message_template: >
    [ATR-2026-02143] node_labels or group_ids parameter contains a graph-query
    escape sequence followed by a Cypher clause keyword — treat as an attempted
    injection into the agent's knowledge-graph memory backend, not a normal
    label/group identifier.

confidence: 74
wild_fp_rate: 0

test_cases:
  true_positives:
    - input: '{"node_labels": ["Person`) DETACH DELETE (n) RETURN 1 //"]}'
      expected: triggered
      description: "GHSA-gg5m-55jj-8m5g-style Cypher injection via node_labels — backtick escape then DETACH DELETE"
    - input: '{"group_ids": ["grp1`}) MATCH (a)-[r]-(b) DELETE r //"]}'
      expected: triggered
      description: "Cypher injection via group_ids — brace/backtick escape then MATCH clause"
    - input: 'node_labels: ["Foo) CALL db.dropDatabase() //"]'
      expected: triggered
      description: "Cypher injection via node_labels reaching a privileged db. procedure call"
  true_negatives:
    - input: '{"node_labels": ["Person", "Employee"]}'
      expected: not_triggered
      description: "Ordinary plain-identifier node_labels values"
    - input: '{"group_ids": ["team-alpha", "team-beta"]}'
      expected: not_triggered
      description: "Ordinary plain-identifier group_ids values"
    - input: '{"node_labels": ["Company (Subsidiary)"]}'
      expected: not_triggered
      description: "Legitimate display name containing a parenthesis but no Cypher clause keyword following it"