Explaining the AI Resilience and Adaptation Framework

On the 6th of December 2025, a seventeen-year-old boy in Japan was arrested for carrying out a cyberattack on the server of Kaikatsu Frontier, an internet café chain operator. The teenager hacked the system by generating code using conversational AI, thereby compromising the data of 7.3 million customers and disrupting all business operations. AI Incident Database reported that the suspect’s prompts to the AI “concealed malicious intent”. The suspect had a case history of unrelated credit card fraud.

The case reveals several facts about the current state of AI. That a teenager was able to access and comprehend powerful AI capabilities easily and cheaply by building sophisticated code, presumably with basic knowledge. Together with the right prompts, he was able to attack a server system that would otherwise need years of expertise to plan and execute.

Clearly, the company’s cyber infrastructure was weak; the AI models are already diffused into our lives. We cannot control who builds it or how people will use it. There will be elements that misuse it, thus exposing the dual capability nature of the models to benefit or harm society.

This pattern isn’t isolated to Japan. In Denmark, a 22-year-old used AI to research how to injure his father without killing him. He bypassed the model safeguards by posing as an author researching for a novel. The AI provided a detailed plan to execute the intended harm.

Let’s look at the magnitude of the teenager’s crime: millions affected by just one action, disrupting business operations all along. How would you protect millions, and in cases where the data exposure results in other harm, such as breaking into the banking systems? How and who would compensate the “millions” of victims?

In the article, I will discuss how adopting the AI resilience framework — which focuses on the societal adaptation through avoidance, defence and remedy interventions — can enhance collective responsibility of Tech companies, governments and society to mitigate AI risks.

AI Resilience complements traditional model-level safety approaches. While the conventional AI safety focuses on training data quality, safeguards, and capability restrictions, resilience addresses what happens when the model-level safety fails or is bypassed, as examples above demonstrate. It is more about how to deal with safety when it is easily accessible (as already is), becoming cheaper to use and deploy for high-end tasks and has a potential for dual use capability. In such a scenario, we need a societal response in terms of adaptation to Advanced AI systems.

Societal Adaptation or AI Resilience begins when AI models are deployed and get diffused. It is at this wide-scale use that new risks are discovered that might slip through even the best of safeguarding measures organically built into the systems, so adaptive interventions are meant to mitigate the harm that might arise in the specific use case of the AI. For example, students use conversational AI for coding or as tutors, but in the above example, one of the teenagers used it for committing a cybercrime.

AI Resilience has a three-part framework:

1. Avoidance: Interventions that stop harmful use before it happens. This includes laws against AI-assisted crimes, age restrictions on accessing harmful content, and monitoring systems that detect suspicious activity. Avoidance aims to prevent attacks from occurring in the first place.

2. Defence — It aims to prevent harm even when the misuse occurs by building systems that withstand it. Interventions include cybersecurity infrastructure that deters cyberattacks, spam filters that catch phishing, and public awareness campaigns to raise awareness of AI harms.

3. Remedy: This intervention reduces the downstream impact after the harm occurs. This includes legal actions such as arrest and prosecution, victim compensation (e.g., people losing money through banking scams) and other rapid responses to contain damage. Remedy faces severe challenges at scale. While legal systems can prosecute one attacker, like in the Denmark case above, they struggle to help millions of victims, like in the Japanese case, where the data of the customers could have been used to hack banking systems.

Let’s deconstruct the Japanese case using the above framework:

Avoidance Failure: In the case of the Japanese teenager, despite his earlier credit card fraud history, the cyber laws didn’t deter him. No monitoring systems tracked his activities, nor any surveillance detected his planning phase, and no age restriction prevented his access to AI’s code-generation capabilities.

Defence Failure: The company’s weak cybersecurity mechanism was vulnerable, and the intrusion succeeded without detection. The generated code succeeded in breaching servers and exposing the data of 7.3 million customers.

Remedy partially worked: Legal accountability led to the arrest of the attacker after months of investigation. However, it compromised the privacy data of 7.3 million people and caused operational havoc which cannot be undone. So one case of misuse of an AI model targeted millions at scale, explaining why remediation systems struggle with AI-enabled mass harm.