Claude Code Strengthens Safety Rules

The Claude Code system prompt and agent rules were adjusted with a focus on security and reporting in 2.1.136.

• The System Prompt now requires confirmation for irreversible or externally impactful actions unless there is continuously valid approval.
• Before deletion or overwriting, it now requires checking the target first, and emphasizes truthfully reporting skipped steps, failed tests, and verified results.
• The Auto mode rule reviewer added hard_deny, so security boundaries are unconditionally blocked, while soft_deny is left only for destructive, irreversible actions that can be approved with clear user intent.
• The Security monitor moved data exfiltration, bypassing safety checks, and inferred external services/download sources to hard-block.
• The Edit tool description restored the line-number prefix template variable, while keeping the existing instruction to exclude line numbers from the edit string.