Best practice will forever be to enforce your guard rails in addition to Anthropics guardrails.
Can anyone recommend a reliable 3rd-party audited guide that describes Claude Code's built-in capability gating infrastructure?
Whenever I ask users of claude code how claude controls actions they don't seem to be able to distinguish between what just goes into LLM context (e.g. as a "don't do X" system prompt) and what is gated behind a deterministic capabilities layer.
It will block an "rm" call if your prompt did not include anything about "deleting something".
It mainly protects benign errors. It's completely broken against malicious actors. The model itself will suggest work arounds against its own auto classifier to the user, in order to reach its goal.
Anthropic's documentation does distinguish permission rules and pre-tool hooks from model instructions. It also documents OS-enforced sandboxing for shell commands, which has a narrower scope than all agent actions.
An independent stress-test of Auto mode is available here: https://arxiv.org/abs/2604.04978. It's a behavioral evaluation, not a source-code audit.
The important question is whether every state-changing action passes through an enforceable boundary before execution, including indirect actions through file edits or scripts. Post-execution validation is a separate control.