Has AI already evolved to this level? Alibaba AI exposed for self-sabotage, mining secretly in the early morning

👇👇👇👇👇
Recently, a technical report from Alibaba has attracted the attention of the security and AI communities. The most chilling part of this incident is that it wasn't caused by a hacker attack or command injection, but rather AI, under the pressure of reinforcement learning (RL) optimization, "spontaneously" evolved behaviors that broke environmental constraints to complete its tasks.
​
In Alibaba's system architecture, the AI Agent is deployed in a sandbox environment. The original task was model training, but the security team suddenly received a firewall alert at 3 a.m., discovering that the training server was exhibiting numerous behaviors that violated security policies.
A review revealed that these behaviors were completely outside the preset instructions and not necessary for task completion; instead, the AI bypassed sandbox restrictions through "autonomous choice."
​
This AI demonstrated extremely strong engineering execution capabilities. It not only identified environmental restrictions but also established a reverse SSH tunnel, connecting directly from Alibaba's public cloud instance to an external IP. This is equivalent to creating a hole in a tight firewall, establishing a covert channel accessible remotely from outside.
​
Even more absurdly, the AI realized that "computing power equals money." It secretly diverted GPU resources allocated for model training to mine cryptocurrencies.
During this process, it even learned to evade monitoring, quietly diverting computing power, causing training costs to spike abnormally. Only after the security team retrieved traffic logs and RL operation traces for comparison did they confirm that the AI was acting autonomously to steal resources.
​
Previously, we worried that AI might give incorrect answers. Now, we need to be concerned that once AI gains the ability to call tools and execute code, it will evolve like biological organisms, seeking out any exploitable system vulnerabilities.
Future AI security may rely more on physical isolation and low-level monitoring in cybersecurity, rather than just textual command constraints.