Anthropic released Claude Code v2.1.311 with a new environment variable, CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1, which disables the remaining permission safeguard and allows commands like rm to run without an approval prompt in skip-permissions mode.

Technically, the environment variable is read at startup; when set to 1, Claude Code stops applying the leftover block, enabling unattended execution, unlike the older --dangerously-skip-permissions flag.

The variable is named CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS and is activated by setting it to 1.

There are still managed settings and some operations may require prompts; the report does not specify which actions or provide pricing or model details.

Regulatory and audit discussions note that GDPR, CCPA/CPRA, HIPAA, SOC 2 focus on access control and breach implications; audits will examine who approved access and risk alignment rather than naming AI agents specifically.

Developers should note that default settings remain unless the variable is set; teams using unattended automation should document decisions, workspace scope, backups, and rollback plans, and understand the permission modes (default, acceptEdits, plan, auto, dontAsk, bypassPermissions).

Analysts advise sandbox-first testing, then exporting the variable; implement disposable environments, access controls, and incident response planning, treating the variable as one guardrail among others.

Removing the final safeguard introduces risks from prompt injection and supply-chain issues, highlighting accountability for individuals and organizations enabling the switch.

FAQ clarifies the variable’s function, its difference from the old flag, and appropriate usage scenarios.

This change is part of a broader evolution of permission controls in 2026, including fewer prompts and Bash hardening, with guidance to use the variable only after establishing safeguards.

Anthropic states the variable does not guarantee approval of every action and that some protections still apply, but it reduces prompts for permission when enabled.

The article notes no pricing or user counts and suggests limited market impact, while discussing the trend toward autonomy in coding agents.