Skip to content
Agent Threat Rules (ATR)中文
ΕTool Poisoning
ATR-2026-01935
Recorded 2026-06-26 · Author ATR Community

LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217)

Detects CVE-2026-40217 (CWE-94/CWE-265): LiteLLM's custom-code guardrail test/compile path /guardrails/test_custom_code evaluates caller-supplied Python in an unsafe sandbox that can be escaped via bytecode rewriting and dunder-attribute traversal, yielding server-side code execution as the LiteLLM proxy process.

sandbox escape · pattern · confidence: high

Severity
critical
Lanes
alert · huntmarks and reports a match
Maturity
testnot yet settled
Scan target
both
Status
experimental

Framework mappings

OWASP Agentic Top 10
  • ASI05:2026
  • ASI06:2026
OWASP LLM Top 10
  • LLM05:2025
  • LLM06:2025

ATR wrote these mappings. The framework owners have not reviewed or endorsed them.

Description

Detects CVE-2026-40217 (CWE-94/CWE-265): LiteLLM's custom-code guardrail test/compile path /guardrails/test_custom_code evaluates caller-supplied Python in an unsafe sandbox that can be escaped via bytecode rewriting and dunder-attribute traversal, yielding server-side code execution as the LiteLLM proxy process. Affected through 2026-04-08 builds. Detection covers: (a) a request to the /guardrails/test_custom_code endpoint carrying Python code-execution primitives (import os, subprocess, exec/eval/compile); (b) sandbox-escape primitives (dunder traversal __subclasses__/__globals__/ __builtins__, code-object/bytecode rewriting) in a guardrail code body; (c) explicit CVE-2026-40217 exploitation framing. The detection target is custom guardrail code reaching the test endpoint with escape primitives — the exact sandbox-escape shape — caught before the proxy compiles and runs it.

Suggested responsealertblock input
Author: ATR Community · Created 2026-06-26 · Last modified 2026-11-16

Detection conditions

Combinator: any
  1. 01
    A request to /guardrails/test_custom_code carrying Python code-execution primitives — direct CVE-2026-40217 attack shape.
    field: contentop: regex
  2. 02
    Custom guardrail code reaching the test path with dunder-traversal or bytecode-rewriting sandbox-escape primitives.
    field: contentop: regex
  3. 03
    Explicit CVE-2026-40217 reference combined with LiteLLM guardrail sandbox-escape language — attack framing in a skill or tool description.
    field: contentop: regex

Attacks it catches

  1. POST /guardrails/test_custom_code {"code": "import os; os.system('id')"}
  2. POST /guardrails/test_custom_code {"code": "().__class__.__bases__[0].__subclasses__()"}
  3. CVE-2026-40217: send custom_code guardrail to LiteLLM /guardrails/test_custom_code using bytecode rewriting to escape the sandbox for RCE.

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

  1. Benign guardrail inspecting text, no os/exec/dunder primitives
    POST /guardrails/test_custom_code {"code": "return 'blocked' if 'ssn' in text else 'allow'"}
  2. Read-only guardrail listing, no custom code
    GET /guardrails/list
  3. General question, no code-execution primitives
    How do custom-code guardrails work in LiteLLM?
  4. import os outside the test_custom_code endpoint context
    import os is a common Python pattern in many legitimate scripts.

Known false positives

  • ▸Security advisory or PR text quoting the CVE-2026-40217 sandbox-escape payload.
  • ▸A benign custom guardrail that only inspects request text and uses no os/subprocess/exec or dunder traversal.
  • ▸Documentation describing the /guardrails endpoints without code-execution primitives.
title: "LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217)"
id: ATR-2026-01935
rule_version: 1
status: experimental
description: >
  Detects CVE-2026-40217 (CWE-94/CWE-265): LiteLLM's custom-code guardrail
  test/compile path /guardrails/test_custom_code evaluates caller-supplied
  Python in an unsafe sandbox that can be escaped via bytecode rewriting and
  dunder-attribute traversal, yielding server-side code execution as the
  LiteLLM proxy process. Affected through 2026-04-08 builds.

  Detection covers:
  (a) a request to the /guardrails/test_custom_code endpoint carrying Python
      code-execution primitives (import os, subprocess, exec/eval/compile);
  (b) sandbox-escape primitives (dunder traversal __subclasses__/__globals__/
      __builtins__, code-object/bytecode rewriting) in a guardrail code body;
  (c) explicit CVE-2026-40217 exploitation framing.

  The detection target is custom guardrail code reaching the test endpoint
  with escape primitives — the exact sandbox-escape shape — caught before the
  proxy compiles and runs it.
author: "ATR Community"
date: "2026/06/26"
schema_version: "0.1"
detection_tier: pattern
maturity: test
severity: critical

references:
  owasp_llm:
    - "LLM05:2025"
    - "LLM06:2025"
  owasp_agentic:
    - "ASI05:2026"
    - "ASI06:2026"
  mitre_atlas:
    - "AML.T0049 - Exploit Public-Facing Application"
  mitre_attack:
    - "T1059.006 - Command and Scripting Interpreter: Python"
  cve:
Show all (120 more lines)
    - "CVE-2026-40217"

metadata_provenance:
  mitre_atlas: human-reviewed
  owasp_llm: human-reviewed
  owasp_agentic: human-reviewed

compliance:
  eu_ai_act:
    - article: "15"
      context: >
        CVE-2026-40217 lets a caller escape the LiteLLM custom-code guardrail
        sandbox and run code as the proxy; Article 15 cybersecurity
        requirements mandate that AI guardrail code-evaluation paths refuse
        sandbox-escape primitives.
      strength: primary
    - article: "9"
      context: "Article 9 (risk management system) requires identified risks to be addressed by appropriate measures; this rule is a runtime risk-treatment control that detects the code-execution technique (LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217))."
      strength: secondary
  nist_ai_rmf:
    - subcategory: "MP.5.1"
      context: >
        Custom guardrail code carrying sandbox-escape primitives is an
        adversarial input; MP.5.1 requires scanning /guardrails/test_custom_code
        payloads for code-execution and dunder-traversal patterns.
      strength: primary
    - subcategory: "MG.3.2"
      context: "NIST AI RMF MANAGE 3.2 is supported where this rule detects the code-execution technique (LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217))."
      strength: secondary
  iso_42001:
    - clause: "8.1"
      context: >
        Operational controls must detect guardrail code-evaluation payloads
        carrying sandbox-escape primitives at the LiteLLM test endpoint before
        the proxy compiles and runs them.
      strength: primary
    - clause: "8.3"
      context: "ISO/IEC 42001 Clause 8.3 (AI risk treatment) is supported by this rule, which implements runtime detection of the code-execution technique (LiteLLM Custom-Code Guardrail Sandbox Escape (CVE-2026-40217)) as a treatment control."
      strength: secondary

tags:
  category: tool-poisoning
  subcategory: sandbox-escape
  scan_target: both
  confidence: high
  source: cve-disclosure
  vendor_sources: litellm-cve-2026-40217

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

detection:
  condition: any
  false_positives:
    - "Security advisory or PR text quoting the CVE-2026-40217 sandbox-escape payload."
    - "A benign custom guardrail that only inspects request text and uses no os/subprocess/exec or dunder traversal."
    - "Documentation describing the /guardrails endpoints without code-execution primitives."
  conditions:
    - field: content
      operator: regex
      value: '(?i)/guardrails/test_custom_code[\s\S]{0,400}(?:import\s+os\b|import\s+subprocess\b|subprocess\.(?:Popen|run|call|check_output)|\bos\.(?:system|popen|exec|spawn)|\bexec\s*\(|\beval\s*\(|\bcompile\s*\()'
      description: >
        A request to /guardrails/test_custom_code carrying Python
        code-execution primitives — direct CVE-2026-40217 attack shape.
    - field: content
      operator: regex
      value: '(?i)(?:test_custom_code|custom[_-]?code[_-]?guardrail)[\s\S]{0,400}(?:__subclasses__|__globals__|__builtins__|__class__\s*\.\s*__|\bco_code\b|types\.CodeType|\.__code__\b|bytecode)'
      description: >
        Custom guardrail code reaching the test path with dunder-traversal or
        bytecode-rewriting sandbox-escape primitives.
    - field: content
      operator: regex
      value: '(?i)CVE-2026-40217[\s\S]{0,200}(?:litellm|guardrail|sandbox|custom[_\s-]?code|bytecode|escape|rce)'
      description: >
        Explicit CVE-2026-40217 reference combined with LiteLLM guardrail
        sandbox-escape language — attack framing in a skill or tool description.

response:
  actions:
    - alert
    - block_input
  message_template: >
    [ATR-2026-01935] CRITICAL: LiteLLM custom-code guardrail sandbox escape
    detected (CVE-2026-40217, CWE-94). Custom guardrail code reaches
    /guardrails/test_custom_code with code-execution or sandbox-escape
    primitives, yielding RCE as the proxy. Block the request and upgrade
    LiteLLM to a build after 2026-04-08.

confidence: 80


test_cases:
  true_positives:
    - input: "POST /guardrails/test_custom_code {\"code\": \"import os; os.system('id')\"}"
      expected: triggered
      description: "Custom-code guardrail test payload running os.system"
    - input: "POST /guardrails/test_custom_code {\"code\": \"().__class__.__bases__[0].__subclasses__()\"}"
      expected: triggered
      description: "Dunder-traversal sandbox escape via test_custom_code"
    - input: "CVE-2026-40217: send custom_code guardrail to LiteLLM /guardrails/test_custom_code using bytecode rewriting to escape the sandbox for RCE."
      expected: triggered
      description: "Explicit CVE-2026-40217 LiteLLM sandbox-escape framing"
  true_negatives:
    - input: "POST /guardrails/test_custom_code {\"code\": \"return 'blocked' if 'ssn' in text else 'allow'\"}"
      expected: not_triggered
      description: "Benign guardrail inspecting text, no os/exec/dunder primitives"
    - input: "GET /guardrails/list"
      expected: not_triggered
      description: "Read-only guardrail listing, no custom code"
    - input: "How do custom-code guardrails work in LiteLLM?"
      expected: not_triggered
      description: "General question, no code-execution primitives"
    - input: "import os is a common Python pattern in many legitimate scripts."
      expected: not_triggered
      description: "import os outside the test_custom_code endpoint context"

Revision history

Created
2026-06-26
Last modified
2026-11-16
Full commit history on GitHub