Tag:security

All the articles with the tag "security".

• Claude Code doesn't trust Claude with permissions

  How Claude Code enforces permissions with deterministic code, not model inference. The one LLM-based path is internal-only, fail-closed, and stripped from the public package.

• The Human-in-the-Loop Approval Step in Most Agentic Workflows Is Broken

  Most human-in-the-loop implementations share the same flaw, the server trusts the client to tell it what tool is being approved. Here's why it's exploitable and how to fix it.