Anthropic’s Claude Code gets ‘safer’ auto mode
ai-safety anthropic autonomous claude
| Source: Mastodon | Original article
Anthropic has rolled out a revised “auto mode” for Claude Code, its AI‑assisted coding assistant, promising to cut down the barrage of permission prompts while tightening safety safeguards. The new feature lets the model decide, within predefined limits, whether to read files, install dependencies or run scripts on a developer’s machine, but it now does so behind a sandbox that isolates execution and logs every action for audit. If a request exceeds the preset risk threshold, Claude Code falls back to the classic approval flow, giving users a clear “middle ground” between full manual control and the earlier, more permissive auto mode.
The change matters because the friction of constant approvals has been a chief complaint among developers who adopted Claude Code after the March 25 launch of its first auto mode (see our coverage on 25 Mar 2026). By embedding real‑time policy checks and a reversible “undo” capability, Anthropic aims to keep the speed advantage of autonomous coding without opening the door to the kind of accidental system changes that have sparked security scares in other AI‑coding tools. Early internal testing, cited by the company, shows a 40 percent reduction in prompt interactions and zero confirmed privilege‑escalation incidents across a pilot of 120 developers.
What to watch next is how quickly the safer auto mode reaches general availability and whether third‑party IDE plugins will adopt the same guardrails. Analysts will also be tracking user‑feedback on false‑positive rejections, which could force Anthropic to fine‑tune its risk thresholds. Finally, competitors such as OpenAI and Microsoft are expected to announce comparable autonomous coding features, setting up a near‑term race to balance developer productivity with robust security controls.
Sources
Back to AIPULSEN