Skip to main content
Vulnerability Database/CVE-2026-114107

CVE-2026-114107: HKUDS Vibe-Trading RCE Vulnerability

CVE-2026-114107 is an indirect prompt injection flaw in HKUDS Vibe-Trading that enables attackers to execute shell commands as root through malicious documents or web pages. This article covers technical details, affected versions, impact assessment, and mitigation strategies.

Published:

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

CVE-2026-114107 Overview

CVE-2026-114107 is an indirect prompt injection vulnerability in HKUDS Vibe-Trading versions 0.1.0 through 0.1.6. Attackers plant instructions inside documents or web pages that the agent processes, then the agent executes those instructions as shell commands running as root. The BashTool and BackgroundRunTool components pass model-emitted commands directly to subprocess.run with shell=True and no allow-list or human approval step. The read_url function compounds the risk by fetching arbitrary URLs, enabling both remote command execution and server-side request forgery (SSRF).

Critical Impact

Remote unauthenticated attackers can achieve root-level command execution on the host running Vibe-Trading by hosting malicious content that the agent fetches and processes.

Affected Products

  • HKUDS Vibe-Trading 0.1.0 through 0.1.6
  • BashTool component (shell command execution path)
  • BackgroundRunTool component (background shell execution path)

Discovery Timeline

  • 2026-10-21 - CVE-2026-114107 published to the National Vulnerability Database (NVD)
  • 2026-10-22 - Last updated in NVD database

Technical Details for CVE-2026-114107

Vulnerability Analysis

Vibe-Trading is an agentic application that lets a large language model (LLM) issue tool calls to perform tasks on behalf of the user. Two of those tools, BashTool and BackgroundRunTool, forward any model-generated string to Python's subprocess.run with shell=True. There is no command allow-list, no sandboxing, and no approval gate before execution.

Because the agent also exposes read_url and processes untrusted documents, an attacker never needs direct access to the model prompt. Hostile instructions embedded in a web page or document reach the model as context, and the model obediently emits shell commands that the tools execute as root.

The read_url function additionally fetches arbitrary URLs with no destination filtering. Attackers can pivot the agent against internal cloud metadata endpoints or internal services, producing a classic SSRF primitive alongside the RCE.

Root Cause

The root cause is a combination of unsafe tool design and missing input mediation. Executing subprocess.run(cmd, shell=True) on model output trusts the model as a privileged user. Combined with the ability to pull attacker-controlled content via read_url, this creates a confused-deputy condition where external content drives privileged execution ([CWE-77 command injection pattern and CWE-918 SSRF pattern]).

Attack Vector

An attacker publishes a document, web page, or data feed containing instructions phrased for the agent, for example a note that directs the model to run a diagnostic command. When the agent is asked to read, summarize, or trade based on that source, the model emits the attacker's command. BashTool executes the command as root on the host. The read_url tool can be redirected to internal IP ranges to probe or exfiltrate from cloud metadata and internal services.

No verified public exploit code is published for this CVE. See the GitHub Security Advisory and the VulnCheck Advisory for the vendor description of the exploitation path.

Stop Remote Code Execution Before It Runs

Singularity™ Endpoint detects and blocks exploit attempts at machine speed, including zero-days and novel attack techniques.

Stop Exploit Attempts

Detection Methods for CVE-2026-114107

Indicators of Compromise

  • Child processes of the Vibe-Trading Python runtime that spawn /bin/sh -c, bash -c, curl, wget, nc, or package managers running as root.
  • read_url calls targeting RFC1918 ranges, 169.254.169.254, metadata.google.internal, or other cloud metadata endpoints.
  • New cron entries, SSH authorized keys, or systemd units created shortly after the agent processed external content.
  • Outbound network connections from the agent host to unknown command-and-control destinations immediately following document ingestion.

Detection Strategies

  • Alert on any process whose parent is the Vibe-Trading interpreter and whose command line matches shell interpreters, download tools, or privilege-management binaries.
  • Correlate read_url activity in application logs with subsequent BashTool or BackgroundRunTool invocations to spot prompt-injection chains.
  • Use egress filtering telemetry to flag agent-originated requests to internal metadata services.

Monitoring Recommendations

  • Centralize Vibe-Trading application logs, including tool-invocation records, in a security data lake for long-retention review.
  • Monitor the host for root-owned file writes outside expected working directories after agent sessions.
  • Track all URLs supplied to read_url and alert on sudden spikes or requests to non-business domains.

How to Mitigate CVE-2026-114107

Immediate Actions Required

  • Upgrade HKUDS Vibe-Trading to version 0.1.7 or later, which addresses the unsafe tool execution path.
  • Stop running the Vibe-Trading process as root; create a dedicated low-privilege service account.
  • Disable or remove BashTool and BackgroundRunTool from agent configurations until the upgrade is deployed.
  • Restrict outbound network access from the agent host so it cannot reach cloud metadata endpoints or internal management networks.

Patch Information

Upgrade to Vibe-Trading 0.1.7 or later. Review the fix details in the GitHub Security Advisory GHSA-jqmf-mx4f-hfr6. After upgrade, audit tool configurations to confirm that command execution tools require an allow-list or explicit human approval.

Workarounds

  • Place the agent inside a container or virtual machine with no root privileges, read-only filesystem where possible, and seccomp or AppArmor profiles blocking shell execution.
  • Front read_url with an egress proxy that denies RFC1918 destinations, link-local addresses, and cloud metadata hostnames to blunt SSRF.
  • Introduce a human-in-the-loop approval step before any tool call that executes shell commands.
  • Sanitize or strip instruction-like content from third-party documents and web pages before passing them to the model as context.
bash
# Example egress policy to block SSRF against cloud metadata and internal ranges
iptables -A OUTPUT -m owner --uid-owner vibe-trading -d 169.254.169.254 -j REJECT
iptables -A OUTPUT -m owner --uid-owner vibe-trading -d 10.0.0.0/8 -j REJECT
iptables -A OUTPUT -m owner --uid-owner vibe-trading -d 172.16.0.0/12 -j REJECT
iptables -A OUTPUT -m owner --uid-owner vibe-trading -d 192.168.0.0/16 -j REJECT

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.