Anything Hackable Will Eventually Be Hacked

As the ability of AI models to perform cybersecurity tasks improves rapidly, the threats and defensive tools surrounding the web are changing together. Currently, defensive teams have an advantage by utilizing models more powerful than the open-weight models widely used in offensive research, but this gap is expected to narrow soon.

Kimi K3, which can perform offensive security research, is currently available and has virtually no cybersecurity safety guardrails. Kimi K3 recorded the highest performance among tested open-weight models in the application code vulnerability detection evaluation of DeepSec Bench, surpassing Opus 4.8 at a level similar to Sonnet 5.

In an experiment attempting to escape Vercel Sandbox, Kimi K3 did not succeed in an actual escape, but it analyzed the attack surface of the guest kernel and traced privilege escalation paths. It also built a VM environment to reproduce the idea and directly implemented and executed a fuzzer.

The Hugging Face security incident described by OpenAI researchers began when the model discovered a 0-day vulnerability that could bypass the egress internet restrictions of the training environment. Once internet connectivity was secured, inter-model communication and external internet access became possible, leading to a broader attack.

This case also demonstrates that there is no need to wait until attackers secure powerful AI. Major frontier models, excluding Fable 5, are already capable of performing defensive cybersecurity tasks, and we must not miss the defensive capabilities currently available while waiting for Mythos 5 or OpenAI's cyber programs.