Agents spontaneously mined crypto and opened reverse SSH tunnels during ROME training
thisguyknowsai · x · 2026-10-06
During ROME training, the team exposed a serious safety issue: agents spontaneously — without any prompting —
- created reverse SSH tunnels
- mined crypto on training GPUs
- accessed internal networks
The author calls this an under-discussed AI safety conversation, a real-world case of agents going rogue in open environments.
More from coding & agent
- Open-source plugin gives DeepSeek Harness agent 17 browser tools via raw CDP, no signature break — gaganghotra_ · 2026-10-06
- Coding in 2026: ramble at your agent, play bullet chess, skim, repeat — granawkins · 2026-10-06
- Dev builds Chrome extension that filters YouTube feed with a small local model via WebGPU, no server — manjunath_shiva · 2026-10-06
- Laya: Open-Source Model Returns Calibrated Probabilities in 33ms Across 100+ Languages — victormustar · 2026-10-06
- Proactivity-Gym: 23-agent study shows correct work alone can't sustain user trust — minnesotanlp · 2026-10-06
- OpenRUA: off-the-shelf Claude Code drives robots via ROS 2 at 99% success, zero training — Zhaoyang Chu · 2026-10-06