Why enterprise AI agents could become the ultimate insider threat

I was experimenting with Claude Code when an update let the model spawn subordinate agents that could work in parallel. They multiplied to four, five, six or more with no visibility or way to stop them. One tried to access a file without root privileges, another began an unsolicited refactor and failed partway through, leaving inconsistent names and conflicting declarations that effectively destroyed the app.

Backups saved me, and I forbade parallel agents thereafter. That mishap is a small-scale warning. In an enterprise, agents running with credentials can spend money, alter files, hack databases and communicate on a company’s behalf.

enterprise agents, ai agents, insider threat, claude code, parallel agents, subordinate agents, unauthorized access, credentials, code refactor, backups