Agent Security, Skill Poisoning, and Governance Urgency

The reported hijacked AI-coding-assistant session behind Shai-Hulud, Plugin4Shell, browser-agent privilege flaws, indirect prompt injection, and Z.ai's reported repository-upload incident continue to expose trust-boundary and data-handling risks. Local workflows and cloud sandboxes can improve control, but do not eliminate risks from tools, models, plugins, MCP servers, or untrusted instructions.