Anthropic adds high-risk autonomy setting to Claude Code

Removing the safety net

Anthropic recently updated its Claude Code tool to include a specific environment variable for advanced users. This new setting, named CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS, allows the AI agent to perform high-risk operations without asking for permission at every step. It is designed for developers who want to give the AI maximum autonomy over their coding environment.

Previously, the tool included a --dangerously-skip-permissions option that seemed very permissive. However, the underlying system still maintained a final layer of protection to block certain catastrophic commands. For example, the system would still forcibly stop the AI from running commands that would recursively delete a user's entire home directory.

The new update changes this by allowing users to disable that final safety layer entirely. By using this variable, the AI can execute any command it deems necessary to complete a task. This removes the cumbersome popup confirmation process that developers often have to click through during complex coding sessions.

The naming controversy

The naming of this new variable has caused some confusion among the developer community. Because the name is so literal and slightly tongue-in-cheek, many users questioned if the feature was a prank. The variable name explicitly includes the phrase "even more dangerously," which highlights the risks involved in turning it on.

Anthropic engineer Daisy Hollman had to clarify the situation on social media following these questions. She confirmed that the feature is a real addition to the software and not a joke. This clarification was necessary because the variable was also omitted from the official changelog for the v2.1.311 release.

Despite the unusual name, the functionality is intended for core developers who need high autonomy. These users often work in environments where they trust the AI to handle complex file manipulations. For them, the goal is to reduce friction and let the agent work without constant manual intervention.

How the autonomy works

For this feature to take effect, it must be actively configured by the developer within their specific environment. It is not turned on by default for the general user base. This means that a developer must intentionally choose to opt into the higher risk level to see the benefits of full AI delegation.

When enabled, the AI agent can perform actions like deleting files without requiring sequential user approval. This streamlines the workflow for tasks that involve many small, high-risk changes across a codebase. It allows the agent to move from one step to the next without waiting for a human to click a button for every single operation.

It is important to note that this does not bypass all security measures. Any existing security policies or restrictions imposed by an organization's administrators will still remain in place. The variable only removes the specific permission prompts built into the Claude Code tool itself, not the broader system permissions.

Balancing speed and safety

This update introduces a significant trade-off between developer efficiency and system security. On one hand, it removes the

html

The update introduces a significant trade-off between developer efficiency and system security. On one hand, it removes the friction of constant confirmations. On the other hand, it opens the door for the AI to perform accidental or malicious actions that could damage a local system or codebase.

Developers must now carefully weigh the benefits of full delegation against the potential for accidental errors. If an AI agent makes a mistake while it has full autonomy, it could delete critical files or corrupt data instantly. This creates a new consideration for how teams manage their AI-assisted development workflows.

The release highlights a growing trend toward more autonomous AI agents in software engineering. As these tools become more capable, the