Vercel argues that compute isolation alone can't stop AI agents from exfiltrating data or probing internal services. Sandboxes need egress control too.
UnYOLO is a credential broker and policy engine for AI agents accessing GitHub. Agents get short-lived proxy tokens instead of real credentials.
A 0-day in Cursor on Windows executes a planted git.exe from any repo root with no user interaction. Mindgard went public with full disclosure after private channels stalled.
A vibe-coded site called Boomberg shipped with a hidden SQL injection vulnerability. The broader pattern is worse than one bad deploy.
Airgap uses Linux namespaces to prevent AI coding agents and npm installs from accessing sensitive secrets like .env files and SSH keys.
Replit published its defense-in-depth security architecture for vibe coding, covering sandbox isolation, supply chain protection, and continuous testing across the stack.
Daniel Stenberg says AI-assisted security reports hit curl at 4-5x the 2024 rate. The quality is higher, the volume is unsustainable, and his wife is worried.
Anthropic published a detailed engineering post on how it sandboxes Claude across Claude.ai, Claude Code, and Cowork — with gVisor, Seatbelt, Bubblewrap, and full VMs.
Cursor launched Security Review in beta for Teams and Enterprise plans, adding two always-on AI agents that scan PRs for vulnerabilities and risky code patterns.
Microsoft published an architecture for securing MCP tool calls with identity, authorization, validation, rate limits, and audit logging. Here's what it means.