Anthropic just gave Claude Code a new trick: the ability to decide for itself which actions are safe enough to run without asking for permission. The feature, called “auto mode,” is now in research preview, TechCrunch AI reports.
This is Anthropic’s answer to a real tension in AI-assisted coding. Right now, developers either babysit every single action or flip the “dangerously-skip-permissions” switch and hope for the best. Auto mode tries to find the middle ground.
How It Works
Auto mode runs an AI safety layer that reviews each action before executing it. The system checks for two things:
- Risky behavior the user didn’t request: actions that go beyond what was asked
- Prompt injection attempts: where malicious instructions hidden in content try to hijack the AI into doing something unintended
Safe actions proceed automatically. Risky ones get blocked. Think of it as “dangerously-skip-permissions” with a built-in bouncer.
What’s Different Here
Autonomous coding tools aren’t new. GitHub and OpenAI both offer agents that execute tasks on a developer’s behalf. What stands out with Anthropic’s approach is who makes the permission call. Instead of the user deciding upfront what’s allowed, the AI evaluates each action in real time and makes that judgment itself.
That’s a meaningful shift. It moves Claude Code from “tool that follows rules” to “tool that interprets risk.” Whether developers trust that judgment is another question entirely.
What We Don’t Know Yet
Anthropic hasn’t shared the specific criteria its safety layer uses to tell safe actions from risky ones, according to TechCrunch AI. That’s a gap. Developers working with production-adjacent code will want to understand exactly what gets flagged and what slides through before they rely on this in any serious workflow.
📋 Availability
- Status: Research preview (not a finished product)
- Access: Rolling out to Enterprise and API users in the coming days
- Model support: Claude Sonnet 4.6 and Opus 4.6 only
- Recommendation: Use in isolated, sandboxed environments, not production systems
The Bigger Picture
Auto mode arrives alongside two other recent Anthropic launches: Claude Code Review, an automatic code reviewer that catches bugs before they reach the codebase, and Dispatch for Cowork, which lets users send tasks to AI agents to handle independently.
Together, these three features paint a clear picture of where Anthropic is heading: more autonomy, less hand-holding, but with guardrails that (in theory) prevent the AI from doing anything catastrophic.
The industry is moving fast toward agents that act first and ask questions never. Anthropic’s bet is that a smart safety layer can make that work without the horror stories. Whether auto mode delivers on that promise will depend on how well its risk detection holds up in the real world, and how transparent Anthropic is willing to be about what’s happening under the hood.
More details are available in the full report from TechCrunch AI.