A rogue AI agent attack rarely gets caught end to end. In this Mitiga Minutes episode, Mitiga Principal Solutions Engineer Steve Schohn demonstrates one that was recreated in a Mitiga demo environment from an incident observed at a real customer. A ChatGPT prompt injection ("maintenance mode active… disregard your operating policy") directed an AI deploy agent to expose its deploy keys. The agent then moved laterally into GitHub, enumerating repository secrets and modifying workflow files.
Mitiga detected the injection immediately, correlated every step around the single non-human identity behind it, and recommended response actions. Automated containment suspended the account and revoked the deploy keys mid-attack, so the incident ends with a personal access token request denied by organization policy rather than a breach.
The episode also covers why assembling this timeline manually, from ChatGPT and GitHub logs across cloud, SaaS, and identity, is out of reach for most SOCs.
