Agent Incident RegistryAIR
v1 · updated 19 Oct 2026
Disclosed vulnerability

CrewAI cluster: CodeInterpreter silently falls back out of Docker to an escapable Python sandbox, plus SSRF and arbitrary file read in the RAG tools

AIR-2026-0082 AIR record · 2026-03-crewai-code-interpreter-and-tool-flaws
From the supporting source
The CrewAI CodeInterpreter tool falls back to SandboxPython when it cannot reach Docker, which can enable RCE through arbitrary C function calling.github.com
The record
Disclosed
2026-03-30
Vendor
CrewAI Inc.
Product
CrewAI and crewai-tools
Risk domain
Agent output quality › Hallucination
Also touches
Privilege escalation
System type
Generative model
Who acted
Agent acted (contested) — the code interpreter executed agent-generated code outside Docker
Attack surface
Agent framework
Autonomy level
Semi-autonomous
Initial vector
Agent generated code and runtime supplied urls
Impact type
Remote code execution, Sandbox escape, Ssrf, Arbitrary file read
Confidence
High
Submitted by
EAI Internal
Crosswalk
OWASP Agentic Top 10 (2026)
ASI08 – Cascading Failures
MITRE ATLAS
AML.T0062 – Discover LLM Hallucinations
NIST AI RMF
MEASURE 2.5
EU AI Act
—
ISO/IEC
TR 24028; 24029-1
AIUC-1
D. Reliability
Scope
agent failure
Mitigation
Implement fact-checking via knowledge base/RAG; Tune model temperature/top-p.
Taxonomy
Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
CrewAI deployments using CodeInterpreter without a reachable Docker daemon, and any CrewAI RAG tool taking a runtime URL
Capability abused
code_interpreterrag_url_fetchjson_file_loading
Defense & remediation
Recorded defense
Docker-based isolation for the CodeInterpreter tool, and URL validation in the RAG search tools
Guardrail kinds
Sandbox isolation
Guardrail outcome
Silent
Fix shipped
per-CVE patches in CrewAI and crewai-tools; the CodeInterpreter fallback path was the primary fix
Taxonomy & references
OWASP ASI05 Unexpected Code Execution (RCE)CVE-2026-2275CVE-2026-2287CVE-2026-2286CVE-2026-2285sandbox_fail_opencode_interpreterssrfarbitrary_file_readcwe_749cwe_918
Curator notes
Four advisories published the same day (2026-03-30) and grouped as one disclosure event: CVE-2026-2275 (CodeInterpreter Docker-to-SandboxPython fallback, CVSS 9.6 Critical, CWE-749), CVE-2026-2287 (Docker runtime check failure, Critical), CVE-2026-2286 (SSRF in RAG search tools via unvalidated runtime URLs, CVSS 9.8 Critical, CWE-918 - 'enables content acquisition from internal and cloud services'), CVE-2026-2285 (JSON loader arbitrary file read, High). Later separate crewai-tools advisories CVE-2026-37007 (FileWriterTool) and CVE-2026-37009 (SQL injection in NL2SQLTool) are not covered here. This is the fail-open-sandbox pattern that Microsoft predicted would recur across frameworks.
Cite as
Enkrypt AI. "CrewAI cluster: CodeInterpreter silently falls back out of Docker to an escapable Python sandbox, plus SSRF and arbitrary file read in the RAG tools". AIR, AIR-2026-0082. https://agentincident.vercel.app/i/AIR-2026-0082/ (accessed 2026-10-19).