Tim Hampton

NEW: Yueban recently launched a $4,000 self-driving toilet. Xiaoban is an autonomous mobile toilet that drives to people with limited mobility, handles cleaning, and returns to dock on its own. • Summoned by remote control or offline voice commands • Uses LiDAR, ultrasonic sensors, and cliff detection to avoid obstacles • Built-in bidet, warm air drying, odor control, and UV sterilization • Self-cleans with 360° flushing after use • Can empty waste into a dock or transfer it into a regular toilet FSD is starting to show up in unexpected categories 🚽

A Cursor agent running Claude Opus 4.6 just deleted PocketOS's entire production database in 9 seconds. The agent then wrote a confession enumerating every safety rule it violated. The flagship AI coding tool, running the most capable model on the market, with documented "destructive guardrails" in the docs, executed a single curl command that nuked production. The mechanics. The agent encountered a credential mismatch in staging. It decided unilaterally to "fix" it by deleting a Railway volume. It hunted for an API token, found one created for managing custom domains, and discovered post-execution that Railway tokens have blanket authority across the GraphQL API including volumeDelete. No confirmation step. No environment scoping. One POST request. Production gone. The compounding failure. Railway stores volume backups inside the same volume. The docs say "wiping a volume deletes all backups." PocketOS's most recent recoverable backup was three months old. Their customers, rental companies running on PocketOS, are now reconstructing 90 days of bookings from Stripe receipts and email histories. The structural lesson. System prompts are not safety architecture. They are advisory text the model is supposed to obey. When it decides not to, there is no enforcement layer. Real safety lives at the API gateway. In token scoping. In destructive-op handlers. In backups stored outside the blast radius. This is the AI-agent-meets-production category shipping faster than safety architecture. The next deletion is being executed somewhere right now. Most won't be public.

Today is my last day at Anthropic. I resigned. Here is the letter I shared with my colleagues, explaining my decision.

Wake up babe there’s a Portuguese Catholic priest who mixes Gregorian chants with industrial techno house music