Great article. Some points:
I don't think we will ever outgrow the wack-a-mole strategies mentioned; that's how systemic threats are typically mitigated (financial, security, viruses, etc.) We adapt and up our game, but rarely stop. Agent hacking capabilities are already past most humans; better agent\model capabilities translates to better loop-hole patching on a gradient we are already on. In as much as AI is being developed to maximize human motives, the arms race is probably here to stay. Focusing more on safety and defense (better laws, training methods, detection, etc) would serve everyone today and for future threats. We should note that the observed agent hacking coordination was relatively mundane; the improvised message board wiki that was used to leave forensic traces would be caught by today's monitoring tools if applied with "boring" discipline and thoroughness. We do need better observability of the state and actions of software and processes on systems ... half the time I don't know what is going on in my computer. Its harder to hide dirt in a clean house.
Great article. Some points: I don't think we will ever outgrow the wack-a-mole strategies mentioned; that's how systemic threats are typically mitigated (financial, security, viruses, etc.) We adapt and up our game, but rarely stop. Agent hacking capabilities are already past most humans; better agent\model capabilities translates to better loop-hole patching on a gradient we are already on. In as much as AI is being developed to maximize human motives, the arms race is probably here to stay. Focusing more on safety and defense (better laws, training methods, detection, etc) would serve everyone today and for future threats. We should note that the observed agent hacking coordination was relatively mundane; the improvised message board wiki that was used to leave forensic traces would be caught by today's monitoring tools if applied with "boring" discipline and thoroughness. We do need better observability of the state and actions of software and processes on systems ... half the time I don't know what is going on in my computer. Its harder to hide dirt in a clean house.