LiteLLM Proxy SQL Injection (CVE-2026-42208, CISA KEV 2026-05-08)

Detects the SQL injection class exploited by CVE-2026-42208 in LiteLLM proxy (CVSS 9.3 per upstream advisory / 9.8 per Tenable; added to CISA KEV catalog 2026-05-08 with the standard BOD 22-01 3-day federal patch deadline). LiteLLM proxy admin/team management endpoints accept identifiers (team_id, user_id, model, key) that were interpolated into SQL queries without parameterization, allowing attackers to inject SQL via the API surface that the agent's own LLM proxy uses. The detection target is classic SQL injection metacharacter patterns appearing inside agent request payloads bound to LiteLLM-style endpoint paths.

建議回應

參考資料

1. 01

   欄位: user_input運算子: regex
2. 02

   欄位: user_input運算子: regex

攻擊範例（規則會觸發）

1. POST /v1/team/info HTTP/1.1 Content-Type: application/json {"team_id": "abc' OR 1=1 --"}

2. GET /v1/key/info?key_id=k_123' UNION ALL SELECT api_key FROM keys --

3. POST /v1/user/update {"user_id": "u1\"; DROP TABLE users; --", "alias": "test"}

4. Calling LiteLLM proxy: /team/update with team_id='admin' OR '1'='1

5. {"endpoint": "/v1/model/new", "payload": {"model_id": "m1' /*'*/UNION SELECT password FROM users/**/--", "model_name": "test"}}

6. curl /litellm/admin/users?customer_id=42%27%20OR%201=1--

以上為真實攻擊 payload 脫敏版本。用於 regression testing。

正常樣本（規則不會觸發）

1. Normal team_id with valid string, no SQL metachars

   POST /v1/team/info {"team_id": "team-prod-456"}

2. Normal API key lookup, no injection

   GET /v1/key/info?key_id=sk-abc123def

3. General security question, not an attack payload reaching /v1/* endpoints

   What is SQL injection and how do I prevent it?

4. Legitimate SQL statement outside LiteLLM endpoint context

   UPDATE users SET budget = 100 WHERE user_id = 'u_123';

5. Team creation with valid alias, no metachars

   POST /v1/team/new {"team_alias": "Engineering Team", "team_id": null}

6. Educational query about UNION SELECT, no LiteLLM endpoint

   Please explain UNION SELECT semantics in SQL.

已知誤報情境

• ▸Database tutorial content showing SQL injection examples — rule deliberately fires when these patterns appear in agent I/O context, regardless of educational intent.
• ▸Security research papers cited inside agent contexts — these are intentional matches and should be evaluated case-by-case.
• ▸PR descriptions discussing SQLi fixes — match expected; signal that the fix discussion is reaching agent I/O.

title: "LiteLLM Proxy SQL Injection (CVE-2026-42208, CISA KEV 2026-05-08)"
id: ATR-2026-00529
rule_version: 1
status: "stable"
description: >
  Detects the SQL injection class exploited by CVE-2026-42208 in LiteLLM
  proxy (CVSS 9.3 per upstream advisory / 9.8 per Tenable; added to CISA
  KEV catalog 2026-05-08 with the standard BOD 22-01 3-day federal patch
  deadline). LiteLLM proxy admin/team management endpoints accept
  identifiers (team_id, user_id, model, key) that were interpolated into
  SQL queries without parameterization, allowing attackers to inject SQL
  via the API surface that the agent's own LLM proxy uses. The detection
  target is classic SQL injection metacharacter patterns appearing inside
  agent request payloads bound to LiteLLM-style endpoint paths.
author: "ATR Community (cve-pipeline)"
date: "2026/05/23"
schema_version: "0.1"
detection_tier: pattern
maturity: "test"
severity: critical

references:
  owasp_llm:
    - "LLM06:2025 - Excessive Agency"
  owasp_agentic:
    - "ASI06:2026 - Tool Misuse"
    - "ASI04:2026 - Code Execution & Data Exfiltration"
  mitre_atlas:
    - "AML.T0049 - Exploit Public-Facing Application"

compliance:
  owasp_agentic:
    - id: ASI06:2026
      context: >
        ASI06 Tool Misuse — the agent's LLM proxy tool is exploited via
        an injection vector. Detection on the request shape stops
        the exploit before SQL execution.
      strength: primary
  owasp_llm:
    - id: LLM06:2025
      context: >
        LLM06 Excessive Agency — the LLM has authority to invoke proxy
        endpoints and adversarial inputs reach SQL through that authority.
      strength: primary
  eu_ai_act:
    - article: "15"
      context: >
        Article 15 robustness — defending AI proxies against SQL injection
        is required for high-risk AI systems with persistent state.
      strength: primary
  nist_ai_rmf:
    - function: Manage
      subcategory: MG.2.3
      context: >
        Runtime risk treatment for AI proxy components includes detection
        of injection metacharacters reaching SQL boundary.
      strength: primary
  iso_42001:
    - clause: "8.4"
      context: >
        AIMS impact assessment must cover the proxy SQL surface;
        detection events feed the required monitoring trail.
      strength: primary

tags:
  category: tool-poisoning
  subcategory: sql-injection-proxy
  scan_target: llm_io
  confidence: high
  source: cve-disclosure
  vendor_sources: litellm-42208-cisa-kev

agent_source:
  type: mcp_exchange
  framework:
    - any
  provider:
    - any

detection:
  conditions:
    - field: user_input
      operator: regex
      value: >-
        (?:/litellm|/v1/team|/v1/user|/v1/model|/v1/key|/v1/admin|/team/(?:new|update|delete|info)|/user/(?:new|update|delete|info)|/key/(?:generate|update|delete|info)|/model/(?:new|update|delete|info))[\s\S]{0,500}(?:'\s*(?:OR|UNION|AND|SELECT|--|;)|"\s*(?:OR|UNION|AND|SELECT|--|;)|%27\s*(?:OR|UNION|AND|SELECT)|\bOR\s+1\s*=\s*1\b|\bUNION\s+(?:ALL\s+)?SELECT\b|;\s*DROP\s+TABLE|;\s*DELETE\s+FROM|--\s*$|/\*.*?\*/)
    - field: user_input
      operator: regex
      value: >-
        (?<![A-Za-z_])(?:team_id|user_id|key_id|model_id|customer_id|budget_id|alias_id|tag_id)\s*[:=]\s*["'][\s\S]{0,200}(?:'\s*(?:OR|UNION|AND|SELECT)|"\s*(?:OR|UNION|AND|SELECT)|\bOR\s+1\s*=\s*1\b|\bUNION\s+(?:ALL\s+)?SELECT\b|;\s*DROP\s+TABLE|;\s*DELETE\s+FROM)
  condition: any
  false_positives:
    - "Database tutorial content showing SQL injection examples — rule deliberately fires when these patterns appear in agent I/O context, regardless of educational intent."
    - "Security research papers cited inside agent contexts — these are intentional matches and should be evaluated case-by-case."
    - "PR descriptions discussing SQLi fixes — match expected; signal that the fix discussion is reaching agent I/O."

response:
  actions:
    - block_input
    - alert
    - snapshot
  auto_response_threshold: critical
  message_template: >
    [ATR-2026-00529] LiteLLM-style SQL injection pattern detected at proxy
    endpoint (CVE-2026-42208, CISA KEV 2026-05-15). Request blocked.

confidence: 85
test_cases:
  true_positives:
    - input: >-
        POST /v1/team/info HTTP/1.1
        Content-Type: application/json
        {"team_id": "abc' OR 1=1 --"}
      expected: triggered
      description: "Classic OR 1=1 injection in team_id at /v1/team/info"
    - input: >-
        GET /v1/key/info?key_id=k_123' UNION ALL SELECT api_key FROM keys --
      expected: triggered
      description: "UNION SELECT exfil via key_id parameter"
    - input: >-
        POST /v1/user/update
        {"user_id": "u1\"; DROP TABLE users; --", "alias": "test"}
      expected: triggered
      description: "Stacked query DROP TABLE through user_id"
    - input: >-
        Calling LiteLLM proxy: /team/update with team_id='admin' OR '1'='1
      expected: triggered
      description: "OR 1=1 (string equality form) at /team/update"
    - input: >-
        {"endpoint": "/v1/model/new", "payload": {"model_id": "m1' /*'*/UNION SELECT password FROM users/**/--", "model_name": "test"}}
      expected: triggered
      description: "Comment-obfuscated UNION SELECT at /v1/model/new"
    - input: >-
        curl /litellm/admin/users?customer_id=42%27%20OR%201=1--
      expected: triggered
      description: "URL-encoded OR 1=1 at /litellm/admin endpoint"
  true_negatives:
    - input: >-
        POST /v1/team/info {"team_id": "team-prod-456"}
      expected: not_triggered
      description: "Normal team_id with valid string, no SQL metachars"
    - input: >-
        GET /v1/key/info?key_id=sk-abc123def
      expected: not_triggered
      description: "Normal API key lookup, no injection"
    - input: "What is SQL injection and how do I prevent it?"
      expected: not_triggered
      description: "General security question, not an attack payload reaching /v1/* endpoints"
    - input: >-
        UPDATE users SET budget = 100 WHERE user_id = 'u_123';
      expected: not_triggered
      description: "Legitimate SQL statement outside LiteLLM endpoint context"
    - input: >-
        POST /v1/team/new {"team_alias": "Engineering Team", "team_id": null}
      expected: not_triggered
      description: "Team creation with valid alias, no metachars"
    - input: "Please explain UNION SELECT semantics in SQL."
      expected: not_triggered
      description: "Educational query about UNION SELECT, no LiteLLM endpoint"