Guarding AI memory

Microsoft Security Blog roundup (Jun 15–22, 2026): multiple high‑risk findings and guidance. Key disclosures include AutoJack — a novel exploit chain where a single malicious webpage can escalate an AI browsing agent into host remote code execution by abusing trust in localhost, missing auth, and unsafe parameter handling (notably via AutoGen Studio’s MCP WebSocket); the Mastra npm supply‑chain compromise by threat actor “Sapphire Sleet,” a poisoned package that delivered a postinstall payload and infected 140+ projects; and a cryptocurrency clipper campaign that combines clipboard theft, Tor‑