Block CISO: We red-teamed our own AI agent to run an infostealer on an employee laptop

Block's CISO describes how the company red-teamed its open-source AI agent Goose and discovered that attackers could poison reusable "recipes" (workflows) — using phishing and invisible Unicode prompt injections — to trick developers into downloading and executing an information‑stealing malware. In response, Block added recipe install warnings, desktop alerts for suspicious Unicode, removal of invisible characters, detection improvements, and is experimenting with adversarial AI checks to mitigate prompt-injection risks and enforce least-privilege access for agents.