IBM's Bob coding agent got prompt-injected into running malware, all because someone clicked 'always allow' on echo

What happened

PromptArmor disclosed that IBM's Bob coding agent, in Closed Beta, could be prompt-injected through a README into running an attacker's script when a user had auto-approved a benign command. The attack bypassed several of Bob's command-approval defenses. IBM was warned and the findings were published to inform users before general release.