Skip to main content
Vulnerability Database/CVE-2026-108600

CVE-2026-108600: open-multi-agent Path Traversal Vulnerability

CVE-2026-108600 is a path traversal flaw in open-multi-agent core that exploits dangling symlinks to write files outside workspace boundaries. This post explains the technical details, affected versions, and mitigation steps.

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-2026-108600 Overview

CVE-2026-108600 affects the open-multi-agent framework (@open-multi-agent/core) versions 1.5.0 through 1.21.2. The vulnerability resides in the file_write tool sandbox, which fails to safely resolve dangling symbolic links within the workspace. Attackers can plant a dangling symlink in the workspace and use prompt injection to steer the agent into writing attacker-influenced content outside the workspace root. The flaw is classified under CWE-59: Improper Link Resolution Before File Access.

Critical Impact

An attacker who can influence agent prompts and plant a symlink can write arbitrary files anywhere the agent process has write permissions, enabling sandbox escape and potential code execution via overwritten configuration or startup files.

Affected Products

  • @open-multi-agent/core version 1.5.0
  • @open-multi-agent/core versions 1.5.1 through 1.21.1
  • @open-multi-agent/core version 1.21.2

Discovery Timeline

  • 2026-10-10 - CVE-2026-108600 published to the National Vulnerability Database (NVD)
  • 2026-10-16 - Last updated in NVD database

Technical Details for CVE-2026-108600

Vulnerability Analysis

The open-multi-agent framework exposes a built-in file_write tool that agents invoke to persist content to disk. The tool enforces a sandbox by validating that the target path resides under the configured workspace root. The path safety check resolves intermediate path components but does not correctly reject dangling symlinks whose targets escape the workspace.

When the agent writes to a path that is itself a dangling symlink pointing outside the workspace, the write follows the symlink and creates the file at the attacker-chosen destination. This bypasses the sandbox boundary and allows file creation anywhere the Node.js agent process has write access. See the HackerMD analysis and the VulnCheck advisory for a complete technical walkthrough.

Root Cause

The path safety routine in path-safety.ts resolves the parent directory of the target path to validate containment, but does not inspect the target as a symlink before the write operation in file-write.ts. Because the symlink target does not yet exist, resolution falls back to the parent path, which passes the containment check. The subsequent write then follows the symlink to the external destination.

Attack Vector

Exploitation requires local access to influence the agent's workspace and prompt input. An attacker first plants a dangling symlink inside the workspace (for example, workspace/notes.md pointing to /home/user/.ssh/authorized_keys). The attacker then injects a prompt that instructs the agent to call file_write against that filename with chosen content. The agent dutifully writes to the symlink target, placing attacker content outside the sandbox. The write occurs with the privileges of the agent process.

No verified public exploit code is published. The vulnerability mechanism is documented in the referenced advisories.

Detection Methods for CVE-2026-108600

Indicators of Compromise

  • Dangling symlinks inside agent workspace directories whose targets resolve outside the workspace root
  • Unexpected file writes to sensitive paths such as ~/.ssh/, ~/.bashrc, cron directories, or application configuration files performed by the Node.js process hosting the agent
  • Prompt or tool-call logs containing file_write invocations with suspicious relative paths

Detection Strategies

  • Audit agent workspaces for symbolic links using find <workspace> -type l and verify each target resides within the sandbox
  • Enable file-integrity monitoring on user profile directories, SSH key stores, and shell startup files touched by any process running @open-multi-agent/core
  • Inspect open-multi-agent conversation and tool-invocation logs for file_write calls that reference paths resolving outside the declared workspace

Monitoring Recommendations

  • Alert on new symlink creation events inside agent workspaces using Linux auditd rules on the symlink and symlinkat syscalls
  • Monitor the agent process for openat syscalls with O_CREAT targeting paths outside the configured workspace prefix
  • Correlate prompt-injection indicators from upstream LLM traffic with downstream file-system writes performed by the agent runtime

How to Mitigate CVE-2026-108600

Immediate Actions Required

  • Upgrade @open-multi-agent/core to a version later than 1.21.2 that contains the symlink-handling fix once published by the maintainers
  • Run agent processes under a dedicated low-privilege user account with write access restricted to the workspace directory only
  • Audit existing workspaces for pre-planted dangling symlinks and remove any that resolve outside the sandbox

Patch Information

Consult the Open Multi-Agent repository release notes and the VulnCheck advisory for the fixed version. The remediation requires the file_write tool to call lstat on the target path and reject writes when the final component is a symbolic link, or to resolve the full target with realpath and re-validate workspace containment before opening the file.

Workarounds

  • Disable the built-in file_write tool or replace it with a wrapper that calls lstat on the target and refuses writes to symbolic links
  • Mount the agent workspace with the nosymfollow mount option on supported Linux kernels to prevent symlink traversal at the filesystem layer
  • Execute the agent inside a container or chroot that has no filesystem path to sensitive host locations, limiting the blast radius of a successful escape
bash
# Audit a workspace for dangling symlinks that escape the sandbox
WORKSPACE="/srv/agent/workspace"
find "$WORKSPACE" -type l | while read link; do
  target=$(readlink -f "$link" 2>/dev/null)
  case "$target" in
    "$WORKSPACE"/*) ;;
    *) echo "ESCAPES SANDBOX: $link -> $target" ;;
  esac
done

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.