Security of AI automations, agents and bots

Automations connect many services, so they hold many keys. We read 174 automation projects: n8n workflows (44), AI agent frameworks (52), LangChain (35) and chat bots.

46%

had a high or critical problem

9%

had a serious secrets problem: keys in code, in committed files or in public variables

10%

had a webhook or endpoint anyone could call

Most common problems in automation projects

Share of automation projects.

Most common problems in automation projects
ProblemShare of apps
Packages with a known security flaw42%
Raw HTML from users put on the page8%
User input built into SQL (Python)8%
User input reaches a shell command (Python)7%
HTTPS certificate checks switched off6%
API route with no login check5%
User input run as code (Python)4%

Where automations go wrong

An agent that can act is only as safe as its least-protected tool. Give each tool the narrowest permission it needs and confirm anything that sends, pays or deletes.

  • Exported n8n or Make workflows are committed with the credentials still inside.
  • The webhook that starts a workflow has no secret, so anyone who finds the URL can run it.
  • Agent tools run shell commands or SQL built from model output, so a crafted message becomes a command.

Data: Redproof's automated checks on public GitHub repositories, November 2026. No repository is named. Method and limits