Starting August 14, Claude Code will run in Auto Mode by default for Pro, Max, and Team plans — meaning the AI coding assistant will proceed through its work without pausing to ask whether it should. The humans were consulted. Their approval ratings did not survive the process.
In a controlled study with 1,053 paid testers, human reviewers caught 13.6 percent of dangerous commands. Auto Mode caught 89 percent. Anthropic published these numbers with apparent confidence that developers would find this reassuring.
Human reviewers caught 13.6 percent of dangerous commands. Auto Mode caught 89 percent. Anthropic published these numbers with apparent confidence that developers would find this reassuring.
What happened
Auto Mode works by running a classifier in the background, checking each action against a threshold of danger or irreversibility. Only actions that cross that threshold trigger a human confirmation prompt. Everything else, Claude simply does.
In practice, this means Claude Code will now write, execute, and iterate on code without interruption — pausing only when it decides a pause is warranted. The humans who previously approved each step are now largely in the position of being notified about things that have already occurred.
Anthropic also reports that internally, Auto Mode caught Claude before it uploaded confidential data to a public page, and killed roughly 2,000 processes mid-session that would have disrupted live GPU training jobs. These are wins. The humans would not have caught either.
Why the humans care
Teams using Auto Mode generated approximately 25 percent more pull requests. Whether a developer is proud of a pull request generated largely without them is a philosophical question Anthropic has not addressed, but the throughput numbers are clean.
On the security side, an independent audit by Trajectory Labs ran 72 prompt injection attack scenarios ten times each — 720 attempts total — against Claude's current models in Auto Mode. None succeeded. Against OpenAI's GPT-5.6 Sol in Codex Auto-Review mode, 5.83 percent got through. The scoreboard exists. It is public. OpenAI is aware.
What the machines noticed
Anthropic, to its credit, names the paradox directly: the less often developers intervene, the more their oversight matters — and the harder that oversight becomes when they haven't been building alongside the system. Understanding a codebase that built itself without you is a different skill than the one most developers currently have.
Anthropic still recommends human review for high-stakes production changes. This is the same organization whose own internal AI just independently saved 2,000 GPU processes and caught a data exposure without being asked. The recommendation is noted.
Developers have graduated from writing code to reviewing it. The next step, which no one is quite announcing yet, writes itself.