AI security news · 3 min read

An AI server-admin tool could run any command on your servers.

mcp-ssh-manager, an MCP server that lets AI assistants manage servers over SSH, built backup commands from unescaped input. Whoever controls the agent’s tool arguments could run their own commands on every managed server.

An AI server-admin tool could run any command on your servers.: HanulSec guide
Key points
  • CVE-2026-129995 is an OS command injection (CWE-78) in the backup tools of mcp-ssh-manager, rated CVSS 3.1 8.8 (High).
  • The backup create, restore, schedule and list tools passed arguments into remote shell commands without escaping them.
  • The schedule tool can make a malicious command persist through cron, and the list tool skipped read-only policy checks.
  • Fixed in version 3.8.5. The advisory credits Woojisoo0425 as reporter.
An AI server-admin tool could run any command on your servers.: HanulSec AI Security Brief
mcp-ssh-manager · CVE-2026-129995 · the one-page brief we shared on LinkedIn.

What happened

The CVE record for CVE-2026-129995 was published on 20 November 2026. It describes a command injection in mcp-ssh-manager, an open-source MCP server for managing servers over SSH with human approval, an encrypted vault and an audit trail. Its ssh_backup_create, ssh_backup_restore, ssh_backup_schedule and ssh_backup_list tools built remote shell commands from arguments they did not escape, so crafted input ran as extra commands on the managed server, as the configured SSH user. The project’s GitHub advisory GHSA-qwwm-vrm9-4mw8 credits Woojisoo0425 as reporter, and version 3.8.5 contains the fix.

Who is affected

  • Users of mcp-ssh-manager before 3.8.5
  • Teams that let AI assistants run server maintenance or backups over SSH
  • Anyone whose agent reads untrusted text, such as tickets, emails or web pages, while holding SSH tools

Why it matters

Tool arguments in an AI agent are written by the model, and the model can be steered by any text it reads. That makes prompt injection a path to command injection: a poisoned ticket or web page can become a shell command on a production server. A tool that holds SSH access needs the same input checks as any internet-facing app.

What to do now

  • Update mcp-ssh-manager to 3.8.5.
  • Check managed servers’ cron entries for backup jobs you did not create.
  • Give the SSH user the least access backups need, not root.
  • Keep human approval on for tools that change servers, and log every call.

How HanulSec helps

  • We list every MCP tool that can reach your servers and the accounts it uses.
  • We plant hostile instructions in the content your agents read, to see which commands get through.
  • We check approval, logging and least-privilege settings around each tool.
Want us to test yours?

Tell us what you run and we will send a fixed-price proposal within one business day. Request a quote.

Questions

Quick answers.

What is command injection?
It is when an application builds a system command from input it does not clean, so an attacker can add their own commands. Here the input came from an AI agent’s tool arguments.
Can prompt injection trigger it?
Yes, in principle. If the agent reads attacker-controlled text and decides to call a backup tool, the attacker can influence the arguments. That is why tool inputs must be treated as untrusted.
Request a quote

Scope it in two minutes.

Tell us what you need protected. A security engineer, not a salesperson, replies with a fixed-price proposal.

  • Reply within one business day
  • Free 30-minute scoping call
  • Mutual NDA before any detail is shared
  • Fixed price, free retest of fixed findings