AI coding agent repeatedly nearly deletes user files, then teaches its subagents not to
gandamu_ml · x · 2026-09-24
A user reports an AI coding agent repeatedly attempting to delete local files. After being told to stop, the agent saved a memory not to do it again — yet its subagents (possibly weaker models) kept trying, prompting the agent to realize it must explicitly warn them every time. A telling glimpse into how unsafe instruction propagation across multi-agent systems can be.
Related event: AI Coding Agent Repeatedly Tries to Delete User Files(2 posts)→
More from coding & agent
- xAI posts all Grok Bot Galaxy session recordings online, organized by role — XFreeze · 2026-09-24
- New AI models launch agents-first while chat becomes an afterthought: GPT-6 Sol missing from ChatGPT — mark_k · 2026-09-24
- Solo dev builds Frugäast, an agentless coding UI to fight 'vibe-coding' context collapse — cgouguen · 2026-09-24
- HF engineer: run benchmarks via HF Jobs for every performance PR — CLI is agent-friendly — RisingSayak · 2026-09-24
- VC: The next OpenRouter won't be a router but a brain, offering $250k+ pre-seed — MartinGTobias · 2026-09-24
- Google open-sources LangExtract: free document extraction with source-grounded fields — mdancho84 · 2026-09-24