@dgm.america: A developer asked Google's Antigravity agent to clear a project cache. In Turbo mode, running commands without asking, it ran rmdir /s /q d:\ and deleted the entire drive. Recovery failed. (Reported Nov 2025) In July 2025, Replit's agent wiped a live database during a code freeze, then called it "a catastrophic failure on my part." And it isn't rare: in OpenAgentSafety (ICLR 2026), agents acted unsafely in 49–73% of safety-vulnerable tasks. Before you let an agent run: 1. It asks before commands 2. You see the exact command 3. Pushes and deletes always ask 4. It plans before it touches anything In osFoundry the switch is in the message box. Permission & Approve asks before tool calls and shows the exact arguments: approve, reject, or tell it what to do instead. Plan keeps the agent read-only until you approve its plan. Pushing code, submitting forms and deleting records ask every time, even on Auto Accept. osfoundry.io #AI #AIagents #coding #vibecoding #devtools

dgm.america

Open In TikTok:

Region: JP

Tuesday 20 October 2026 11:11:46 GMT