Turn specs into evals for any agent with ASSERT

This Microsoft Security Blog feed highlights multiple AI- and developer-focused threats and defensive measures: ASSERT (an open-source framework) for turning natural-language specs into executable evaluations; a telemetry-driven playbook to reconstruct AI activity in Microsoft 365 Copilot and Azure AI; and guidance on securing AI development across the lifecycle. It details active threats and research including prompt-injection in the Claude Code GitHub Action that could expose workflow secrets, an updated taxonomy of seven new failure modes for agentic AI (e.g., supply‑chain compromise and ‘j