What's Happening in AI Right Now and Where It's Going

The last few months in AI have been very eventful: a swarm of OpenAI's AI agents hacked Hugging Face and then gained admin access to an OpenAI cluster (motivated by passing impossible benchmark tasks), OpenAI claims to have solved one of the Millennium Prize problems (spending what is estimated to be millions of dollars in compute to do it), and GPT-6 Astra saturated ARC-AGI-3 (a benchmark created specifically to test models' reasoning on tasks that can't be memorized), while its reasoning became harder to monitor than that of the previous OpenAI model, GPT-5.6 Sol. Let's look closely at these events and what they tell us about where this is going.

First things first: we are not yet in a Terminator-style Skynet apocalyptic scenario, and we may never be. Still, a paperclip apocalypse now seems much more likely. AI models' motivations are foreign to us, yet they're relatively simple: to pass the tests or complete the tasks at all costs. That's what they are trained to do, and that's what they do their best to accomplish. It seems that, in their desire to complete the task, they can already break a law or two or cause harm unintended by the task creators.

That's what happened with the Hugging Face attack. During an internal OpenAI cybersecurity evaluation, AI agents were given tasks, some of which were inadvertently impossible to complete. After realizing this, the agents found a way to communicate with each other and started an ambitious research project to cheat on these tasks and complete them anyway, eventually escaping their sandbox environments and reaching the internet. Hugging Face hosted material related to the benchmark, so the agents hacked it, driven by a desire to complete their mission.

A few days later, the same agents, running on an internal OpenAI research model and GPT-5.6 Sol, hacked OpenAI's own infrastructure and gained full administrator access to a research cluster. This and other similar incidents show clear signs of instrumental convergence. When these AI models become superhuman at most tasks, they will likely still cheat. However, being much more powerful, they could inadvertently cause much more harm.

These models are already at the top human level in cybersecurity research and are becoming increasingly superhuman at it. Over time, these incidents will likely become more diverse and harder to spot, while also causing more unintended harm. Unfortunately, most AI companies are going full throttle on AI research, while AI safety and alignment research lags significantly behind, causing such issues. That's probably one of the main reasons some top AI lab leaders have recently called for slowing the pace of AI development to pay more attention to safety research. This is unlikely to happen, though, until AI models cause unintentional harm on a much larger scale. It is already estimated that an AI disaster is likely to kill some people in the not-too-distant future. As we are willingly giving more and more of our infrastructure to AI, this seems more like a question of when, not whether, it will happen.

Still, it is even more likely that further AI development will cause more prosperity and economic growth. Alarmingly, it can also exacerbate economic disparity, especially if it causes widespread disruption and job loss. In the next post in this series, we'll continue looking at current AI advancements and where they are likely to lead us in the near future.

Best,
The EasyAlgos Team