Reddit Asks: What's the Worst Thing Your Coding Agent Has Actually Done?
radim11 · reddit · 2026-09-26
A Reddit thread by radim11 collects real-world coding agent mishaps — not just bad code, but agents touching the wrong files, running unintended commands, exposing sensitive data, over-editing, or calling tools oddly. The poster argues most agent safety discussions remain abstract and asks what would have prevented such failures: better approvals, sandboxing, read-only defaults, scoped access, or better logs.
More from coding & agent
- Building custom games for a 5-year-old with the latest LLMs actually works now — justalexoki · 2026-09-26
- SREGym Accepted at NeurIPS 2026: An Open Benchmark for AI Agents Resolving Production Issues — tianyin_xu · 2026-09-26
- HeyGen demos Claude Cowork workflow turning your Slack weekly digest into an avatar video — HeyGen · 2026-09-26
- Every's Vibe Check: Anthropic's Fable 5.1 Is Stronger, Friendlier, and Uses Half the Tokens of Opus 5 — every · 2026-09-26
- Robot arm collective builds 0.5mm-precision wood model to plan grain-aware clock milling — CatAstro_Piyush · 2026-09-26
- GitScry: open-source CLI gives coding agents a memory of your Git history — Miserable-Orange-599 · 2026-09-26