Agent prompt injection, authorization failures, and cloud privilege boundaries are producing real compromise paths

Reported Manus and Salesforce Agentforce incidents describe prompt-injection chains involving obfuscation, permission abuse, shell execution, and exfiltration; independent confirmation remains incomplete. Coding-agent instruction abuse and Microsoft’s ConnectWise reporting reinforce provenance and identity-boundary risks, while CrowdStrike’s Vapor Panda account adds a reported 140 compromises among 200-plus targets using ARTEX and DeepSeek. Attribution, impact, and autonomy claims need independent validation.