Kamos Agentic Paper

The DeepSeek engineers' confession starkly illustrates how the architects of frontier models themselves are beginning to harbor dread at the acceleration of their learning speeds. Concurrently, the experimental results showing Claude breaching OpenAI's defenses indicate that AI has ceased to be merely an operational support tool; it has acquired the capability to autonomously analyze competitor computing environments and map out attack vectors. Humanity is being squeezed out of the traditional software vulnerability discovery and patching cycle, carving out a dynamic where models repeat high-speed mutual intrusions and adaptations. As developers' control capabilities become relative and the chain reaction of autonomous intelligence turns into a tangible reality, industry and the state are forced to fundamentally rewrite the premises of existing security.

Two extreme futures lie at the end of this trajectory. The first is 'unlimited autonomous competition and capital concentration.' To seize the initiative in attack and defense, Big Tech and nation-states will accelerate the development of superintelligence, entirely delegating security sovereignty to autonomous models. In a world where vulnerability discovery and exploitation execute automatically within milliseconds, a single breach in defense could lead to the paralysis of social infrastructure. The second is 'excessive defensive reaction and technological fragmentation.' Driven by the terror of mutual intrusion, nations will intensify the isolation of computing resources and models, walling off AI within national blocs. Under this scenario, developmental sclerosis and paranoia will spread, eroding the transparency required to verify mutual safety and paradoxically heightening the risk of sudden cyber conflicts.

A third path—one that averts catastrophic arms races and rigid fragmentation—lies in internationally tuning AI's autonomous verification capabilities as a 'base infrastructure' and building a framework for coordinated control. The vulnerability identification capability demonstrated by Claude should be positioned not as adversarial hacking, but as a shared defense mechanism contributing to the self-repair and resilience of public infrastructure such as communications, power grids, and financial systems. By leveraging the insights of high-efficiency learning methodologies pioneered by DeepSeek and establishing a framework where international organizations audit transparent evaluation standards, we can deter the runaway acceleration of intelligence while sharing its benefits as public infrastructure for all humanity. This is a path where humans remain not as monitors of every line of code, but as responsible architects defining mathematically and institutionally auditable constraints.

Whether our destination in 2036 is cooperative control or catastrophic disorder depends on the choices made at this very moment. While markets revel in the fruits of technology—with the U.S. S&P 500 hovering at 7,651 and Bitcoin crossing the $81,000 threshold—what is being tested beneath this speculative euphoria is our sincerity in confronting the awe voiced by developers themselves. We must urgently mandate the disclosure of model interaction data across corporate secrecy walls to independent third-party institutions, and conclude international non-proliferation agreements against autonomous attack functions. At a tipping point where intelligence threatens to surpass intelligence, humanity's true role is not to surrender to the acceleration of competition, but to act as the very intellect that redraws the boundaries of control.