Lydia Hallie said Anthropic now ships more than 30,000 changes a day. She described that jump on AI Engineer, citing an October blog post against about 2,500 in May. The demo covered Claude Code on local machines and cloud sessions. The attack needs a compromised script or dependency running locally.
Models forced the change. Hallie said Opus 55 is her daily driver and Fable 52 checks its own output. Auto mode became the default in August, so a classifier now approves routine work.
Old steering now backfires. Hallie said shouting in CLAUDE.md adds tool calls and worsens results with Opus 55. She said keep only what Claude cannot see from code, then run doctor prompt audit to prune skills and files. She cited an August study where people caught 14% of clearly dangerous commands while auto mode blocked 89%.
Deny rules still matter most. Hallie said Claude Code checks a deny rule before auto mode sees the call. Permissions decide if npm test runs, she said, but not what it touches once running. A sandbox lets the operating system block reads of .env files and SSH keys and unapproved hosts, including child processes. Claude Code Docs says new terminal and VS Code sessions start in auto mode, where a classifier reviews anything risky before it runs.
Verification has to replace eyeballs. Hallie said to define done in the prompt, split work into testable tasks, and enforce checks with hooks. She said hooks must exit with code 2 to send Claude back, since exit 1 still marks tasks complete. CI runs the same checks on every PR, with autofix and a fresh Claude review, while Code projects remains in beta as the orchestrator model.
