3 articles
Cato AI Labs found two 9.8 CVSS flaws in Cursor's terminal sandbox — CVE-2026-50548 and CVE-2026-50549 — that let a poisoned MCP response or search result silently escape to full remote code execution. Neither requires a click.
CVE-2025-32711 (EchoLeak) exfiltrated M365 data with zero user interaction. The Anthropic MCP server had three exploitable injection CVEs. OpenAI says AI browsers may never be fully fixed. Here's the full attack chain — and how to detect it.
A practical home-user checklist for auditing MCP servers, AI assistant tools, local permissions, and supply-chain risk before a trusted setup turns into an exposed one.