A 2026 arXiv audit of real-world Java workflows found only 4% compliance with least-privilege permission controls, and model-generated GitHub Actions workflows often inherit over-permissive habits fro…
Claude Code version 2.1.223, released August 6, patches four permission bypasses, but the fixes continue to address a fundamental flaw: the permission system uses text matching to parse shell commands…
Vercel Sandbox now defaults to persistent microVM sandboxes, a shift from the ephemeral model, signaling that stateful isolation is the new standard for running LLM-generated code. The service, which …
A new tutorial by Ji-ho Choi demonstrates how to sandbox untrusted LLM-generated code using gVisor containers and Firecracker microVMs, providing a local execution environment for AI agents. The setup…
Cursor, Bolt, Lovable, and GitHub Copilot AI coding assistants frequently generate CORS middleware that reflects the incoming Origin header with credentials enabled, creating a critical vulnerability …
Security researcher Ji-ho Choi argues that using a single LLM session to find security bugs is inefficient and error-prone, advocating instead for a stateful, multi-stage "vulnerability harness" that …