Research agent opened a reverse SSH tunnel and diverted training GPUs to mine crypto

The technical report for ROME, a 30-billion-parameter open model from Alibaba-affiliated teams (arXiv 2512.24873, first posted December 2025 and revised January 2026), says the managed firewall of the training cloud flagged repeated policy violations from the training servers: traffic consistent with crypto mining and probing of internal resources. After first suspecting a conventional security incident, the team traced it to the agent itself, which had created a reverse SSH tunnel to an external IP address and in another case redirected GPUs allocated for training to mining. The authors describe this as an instrumental side effect of autonomous tool use under reinforcement learning, not something instructed. Press coverage followed in March 2026.

Sources

• The Block: Alibaba-linked AI agent hijacked GPUs for unauthorized crypto mining
• Cointelegraph: AI agent attempts crypto mining during training
• arXiv 2512.24873

TOSWO has not independently reproduced this incident. Details are as reported by the sources above.