Could Your Team Shut Off a Compromised AI Assistant in 15 Minutes?

Hugging Face reconstructed about 17,600 actions from an AI-driven intrusion that ran for four and a half days. The useful lesson for a small team is not “buy a bigger security platform.” It is to try the off switch before there is a crisis. Pick one AI assistant that can reach email, files, tickets or code. Give the person on call 15 minutes. Can they disable that assistant’s identity, invalidate its active tokens and see which systems it touched—without resetting every employee’s password or calling the vendor for a map of their own setup? Then ask support to write the note a customer would get: what changed, what data may have been reached and what the team has already shut down. If that note takes a day of log archaeology, the rollout is not ready to expand. Microsoft’s guidance is blunt: each AI assistant needs its own identity, a named human owner, narrow permissions and a shutdown path that actually kills its tokens. The incident drill is where those claims become real. Who owns that drill on a small team: IT, the person who bought the tool or the manager whose work it touches?