Adversa AI June 2026 Agentic Security Roundup: IterInject Optimizes Prompt Injections Iteratively, Semantic Kernel RCE, Copilot Backdoor

Adversa AI·medium signal

Adversa AI's June roundup highlights three new threats: IterInject, a technique that optimizes indirect prompt-injection payloads iteratively using model feedback instead of handcrafted strings; a Microsoft Semantic Kernel flaw chaining prompt injections to host-level RCE; and a DEF CON demonstration of a persistent Microsoft Copilot backdoor via indirect injection.