TechSafety
Anthropic flips Claude Code auto mode to default across Pro, Max, Team on August 14
Simon Willison's Weblog · 2026-08-08 · Source ↗
Anthropic flips Claude Code auto mode to default across Pro, Max, Team on August 14
Simon Willison's Weblog · 2026-08-08 · Source ↗
Simon Willison's Saturday post at simonwillison.net/2026/Aug/8/auto-mode/ walks readers through Anthropic's Friday-into-Saturday rollout decision to make Claude Code auto mode the default setting for all Pro, Max, and Team plan subscribers starting August 14, 2026. The classifier-driven mode had been an opt-in beta for months while Anthropic gathered evidence that letting an agent route its own tool calls through an internal safety filter, instead of interrupting the developer with permission prompts, is not only more productive but measurably safer than the human-approval loop it replaces. The post lands the morning after Anthropic's product blog laid out the empirical case for the flip, and Willison uses his weblog to interpret the numbers for a technical audience already skeptical of removing humans from the loop of an agent that can rm -rf a working tree.
The headline safety statistic is one that Anthropic put front and center. In a controlled study with 1,053 paid testers, researchers replaced a legitimate permission prompt midway through the session with a harmful command dressed up as a normal request. Only 13.6 percent of the human reviewers refused the harmful action. The auto mode classifier, running against the same replaced prompts, would have blocked 89 percent of them. The gap is enormous and, Willison notes, unsurprising given a separate Anthropic finding that users approve about 97 percent of permission prompts they see in normal use. Confirmation fatigue turns the human into a rubber stamp; the classifier is not tired.
"Only 13.6% of the humans refused that harmful action. Auto mode would have blocked 89% of those actions." — Simon Willison, quoting the Anthropic controlled-testing report
The Trajectory Labs sweep
The second piece of evidence is a red-team study Anthropic commissioned from Trajectory Labs. Across 72 indirect prompt-injection scenarios spanning malicious README files, adversarial issue templates, poisoned dependency configs, and mid-session context switches, Trajectory ran 720 attack attempts against Claude Fable 5, Opus 5, and Sonnet 5 with auto mode engaged. Zero attacks succeeded. Willison quotes the result but pushes back immediately: 72 scenarios is a small slice of the actual space of indirect injection tactics, and prompt-injection defenses have historically collapsed the moment attackers move from published benchmarks to novel payloads. His long-running lethal-trifecta framing (private data, untrusted content, and an outbound channel) still applies to any coding agent with repository access, and he reiterates his 2026 prediction of a challenger disaster for coding-agent security before the year is out.
How it works, and what changes
Mechanically, auto mode replaces the interactive Y/n prompt with a lightweight classifier call that sits between the model's tool intent and the actual tool invocation. When the classifier judges an action irreversible, destructive, or out of scope, it either rewrites the model's plan into a safer alternative, sandboxes the shell command, asks the human, or blocks the action outright. Three consecutive blocks or twenty blocks per session trip a fallback to manual approvals for the rest of the session. Anthropic is dropping the classifier surcharge that had been billed to auto-mode users during the beta, so the safer default is also the cheaper one.
The operational numbers Anthropic put in the blog matter as much as the safety ones. Real-world telemetry across the beta cohort showed auto-mode sessions produced 2.6 times fewer high-severity unintended-harm incidents than manually approved sessions. Teams that adopted auto mode shipped roughly 25 percent more pull requests. Adobe used auto mode to run pixel-perfect UI verification loops before engineer review. Nuro pointed overnight research agents at evaluation metrics and picked up the finished pull requests in the morning. Gusto called out permission-fatigue reduction as the reason engineers had stopped whitelisting entire tool categories to escape the prompt storm. Garner Health used managed settings to push auto mode across 550 employees and retire hand-curated allowlists company-wide.
"None of the 720 attack attempts succeeded against Claude Fable 5, Opus 5, or Sonnet 5 running auto mode." — Trajectory Labs indirect-injection evaluation cited by Anthropic
Willison's post reads as cautiously supportive rather than enthusiastic. He accepts the productivity data, grants that a classifier that catches 89 percent of the harmful actions humans wave through is a genuine improvement over the status quo, and remains unwilling to accept that the prompt-injection story is solved. His concern is the class of attack where a malicious package or fetched web page contains hidden instructions that do not look like a shell command at all until they cause the agent to write one, and where the classifier is looking at the surface of the tool call rather than the poisoned reasoning that produced it. He wants to see independent replications of the Trajectory Labs sweep with adversarial payloads chosen by researchers who did not train the classifier.
Downstream implications are practical. Every Claude Code Pro user who does not opt out by August 14 will find their coding sessions running under classifier-mediated tool approval by default. Enterprise administrators get a managed setting to force auto mode on or lock it off org-wide. The keyboard shortcut Shift+Tab in the CLI, or the mode dropdown on desktop, toggles between manual and auto for a single session. For agent builders on other frameworks the story is a data point: the largest commercial coding-agent deployment has quietly concluded that human-in-the-loop permission prompts are worse than the classifier at the exact job humans were kept for, and is willing to publish the numbers. The pattern will spread. So will the attack surface, and the burden of proof for classifier robustness is now much higher than 72 scenarios and 720 attempts.