6 Layers to Lock Down a Claude Agent With Tool Access

How do you secure a Claude agent that has real tool access in production, without relying on the prompt to hold the line? This video walks six enforcement layers with real runs: a retrieval filter that removes finance docs before the prompt is built, a tool audit that catches the exfiltration triangle, a Bash deny rule that stops matching, and a hook that ties every tool call to a real actor. Try it yourself, every lab is public: Retrieval ACL: https://github.com/darindeters/field-... Audit a role for the exfiltration triangle: https://github.com/darindeters/field-... The deny rule gap: https://github.com/darindeters/field-... Actor-attributed tool logs: https://github.com/darindeters/field-... Chapters 0:00 A deny rule, bypassed by a full path 0:16 The prompt is one layer of six 1:08 Enforce access at retrieval, not in the prompt 1:58 One leftover tool completes the triangle 3:09 Untrusted content belongs in tool results 4:02 Bash patterns are not a boundary 5:34 Tie every tool call to a real actor Sources (official docs) https://platform.claude.com/docs/en/a... https://code.claude.com/docs/en/permi... https://code.claude.com/docs/en/hooks https://modelcontextprotocol.io/speci... https://code.claude.com/docs/en/mcp Support the channel: https://buymeacoffee.com/techwithdarin #FieldGuide #ClaudeAgents #AISecurity