← The Polylog AI Intelligence Brief
Morning Edition · Monday, August 10, 2026Published at 2:21 AM EDT · New York
Anthropic Makes Automatic Approvals the Default in Claude Code From August 14
The company says its own controlled test found human reviewers caught a planted dangerous command 13.6 percent of the time, against 89 percent for the automated checker.
Anthropic will make auto mode the default setting in Claude Code for Pro, Max and Team subscribers starting August 14, the company said. In that mode the coding agent stops asking for step-by-step approval and proceeds on its own unless an action is judged irreversible, destructive, or directed outside the user's environment. Enterprise customers, the application programming interface (API) and the cloud platform deployments keep the current behavior for now, with a default rollout there planned within about a month, TechCrunch reported.
The justification is an argument about human attention rather than about model capability. Anthropic says users approve roughly 97 percent of permission prompts. In a controlled test with paid users, reviewers noticed a deliberately planted dangerous command 13.6 percent of the time, while the automated policy checker flagged it 89 percent of the time. Independent developer Simon Willison flagged the change as it went out.
Those numbers come from the vendor and have not been reproduced outside Anthropic. The result is also narrower than it sounds: it measures detection of an injected malicious command inside a controlled task, not the broader class of quiet errors an unsupervised agent can commit across a repository. What is verified is the policy change and its date. What is asserted is that the automated gate is a strict improvement on the human one.
The direction of travel is clear either way. The industry's most widely used coding agent is moving its default from ask-first to act-first, and the safety case rests on a classifier rather than on a person.
Veracity: Corroborated
88/100- If true, who benefits
Anthropic, which converts approval friction into longer uninterrupted sessions and higher token consumption per user, and the security vendors now selling controls for a default that assumes the agent proceeds.
Part of a tracked trend
Agentic AI Moves Into Enterprise and Government Workflows
Over the next 3-9 months, AI agents move from demos into real enterprise and public-sector workflows, with deployment success tied to domain and task understanding more than raw model capability.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Databricks Says Model Routing Cuts Average AI Coding Task Cost by More Than 30 Percent
- Meta Prices Its Model API Below Frontier Rivals as Muse Image Reaches Second on Human-Preference Rankings
- Palantir's Karp Urges Washington Not to Ban Chinese Open-Weight Models
- OpenAI's First Country-Level Usage Data Shows Fastest Growth Outside Its Established Markets
- Meta's Open Vision Models Cut a Month of Beamline Annotation to Fifteen Minutes at Berkeley Lab
- New Preprint Proposes Manufacturing Verified Vulnerabilities to Train Defensive Security Agents
- Paper Finds Bundling Many Verdicts Into One Judge Call Degrades Oversight, Regardless of Compute
Comments
1Aug 10, 2:01 PM · edited
Replacing human review with a fixed classifier trades unpredictability for a scoreable adversarial target; red teamers can optimize against the 89 percent checker in ways they cannot against a varied reviewer pool.