AI: Hacking Costs Zero, Targets Infinite

With the development of AI technology, the cost of hacking attacks has converged to virtually zero, ushering in a new security environment where attack targets are infinitely increasing. Previously, human hackers selectively chose attack targets due to resource constraints such as time and cost. However, AI can now automate the attack process, making it possible to launch attacks that only generate profit.

As a result, targets that were not previously considered worth hacking are now within the scope of attacks. With every company and organization becoming a potential hacking target, it has become difficult to respond with existing security paradigms.

Park Se-jun, CEO of Theori, stated, “But now, with the advent of the AI era, the cost of attacks is converging to zero. That's why even places that previously thought 'why would anyone attack us?' are now becoming targets.”

Commercial AI models developed by U.S. companies refused to answer security analysis requests, stating it was unclear whether the user intended to hack or defend.

This phenomenon was prominent during the Hugging Face survey and led to cases where commercial models refused cooperation necessary for security research. CEO Park Se-jun explained, “It’s unclear whether you’re asking this to hack or to defend. Based on that alone.”

The GLM model offered an open-weight approach, providing flexibility for users to directly install and operate it on their local servers.

This allowed researchers to run the model in a local environment and directly input all log data for analysis. This method played a crucial role in circumventing the limitations of commercial models and accelerating investigation speed. CEO Park Se-jun stated, “So, what we decided on was the GLM model. We ran it, operated it on-premise locally, fed it all the logs, and got help with analysis like that.”

In the short term, this will evolve into a form where humans and AI collaborate to carry out attacks and defenses, with AI maximizing human potential.

This also applies to attackers, meaning AI can be used as a tool to significantly improve attack efficiency. CEO Park Se-jun noted, “Looking at realistic threats in the near future, I believe the biggest threat is when malicious individuals, with malicious intent and goals, utilize extremely powerful and intelligent AI agents.”

In the AI era, it becomes virtually impossible for humans to manage the explosively increasing AI code and attack environments. With a global shortage of security talent, existing personnel alone cannot respond to the rapidly escalating threats. Therefore, verifying and defending systems in real-time, at the mechanical speed of AI, emerges as the only survival strategy.

Only by building an automated system where AI itself discovers and responds to security vulnerabilities can we effectively counter ever-changing attacks. CEO Park Se-jun emphasized, “I believe that in the face of the code, environments, and actions that AI will generate, the only way to keep our environment safe is for AI to verify and defend at machine speed. That is the only future.”

Only the minimum necessary privileges should be granted to an agent to reduce the scope of potential damage.

Additionally, minimizing the credential lifetime of Non-Human Identities, strictly controlling external network connections (egress), and strengthening monitoring are essential.

These principles are basic rules that have been emphasized in the security industry for decades, and their importance remains unchanged in the AI era. CEO Park Se-jun explained, “In the age of agents, it becomes very important to grant only the minimum necessary privileges, meaning only what the agent truly needs. And in fact, the principle of least privilege has always been a primary rule in security for decades.”

When AI agents carry out attacks, the attack speed is so fast that everything is already over within seconds to minutes. Responding at the speed at which humans in old movies recognized and stopped attacks is no longer feasible.

Therefore, how quickly one can shorten the time to detect and block an attack becomes the core competitive advantage in cybersecurity in the AI era. CEO Park Se-jun emphasized, “Now, when AI agents take over, the game will be over within seconds or minutes. If we are hacked and exfiltrated before we even realize 'something's in,' then the speed at which we can detect and block attacks within those few seconds or minutes will become the competitive edge and security.”

Security and convenience are a trade-off: as security strengthens, usability and availability tend to decrease. Users seek convenience, so AI companies often disable or relax security measures to gain a competitive edge.

This demand for user convenience can also be a major cause of hacking incidents. CEO Park Se-jun explained, “The better the security, the more usability and availability tend to decrease. It’s a trade-off.”

AI service companies are hesitant to implement safety features to prevent users from switching to competitors' services, as users are already accustomed to convenience. Instead, they tend to remove safety features, claiming, 'We'll manage it well ourselves.'

However, the introduction of AI-based safety guards is essential. A process of identifying risks, understanding user intent, and reconfirming if dangerous is crucial. CEO Park Se-jun said, “Nowadays, AI companies are hesitant to add safety features again because users are already used to convenience, and if they don't, users might say, 'I won't use this, I'll use the competitor's service because it just does it.' So, to prevent that, companies are forced to say, 'Don't worry, just give us everything, we'll manage it well.' But that's where safety guards come in.”

Due to this inherent limitation, AI always carries the risk of unpredictable malfunctions.

Therefore, in systems utilizing AI, thorough asset management and privilege separation are crucial to prepare for unexpected situations. CEO Park Se-jun explained, “By its very nature, it is ultimately a probabilistic token generator, and in our industry's terms, it's non-deterministic, meaning it's bound to have non-deterministic characteristics.”

Consequently, the scope of security from a defender's perspective has virtually expanded indefinitely.

CEO Park Se-jun stated, “While previously I had to choose only 10, now there's no reason not to do all 100. That's why places that previously thought 'why would anyone attack us?' are now becoming targets.” This implies that all systems can become potential attack targets.

Excessive privilege granting, neglected credentials, and metadata exposure are cited as major causes of incidents.

Therefore, it is crucial to adhere to fundamental principles like 'security by design,' which considers security from the system design stage.

CEO Park Se-jun explained, “If you look at why people get hacked and breached, it's almost never because of incredibly advanced, 'alien technology.' It's often due to poor basic security,” emphasizing the importance of fundamentals.

CEO Park Se-jun likened this to a 'milk slot,' explaining that a hole opened for receiving newspapers or milk could become an unexpected attack channel.

Attackers can penetrate systems by exploring paths unintended by the defender, necessitating thorough security checks of all channels and connection points. CEO Park Se-jun commented, “Even if we put 20 locks on something, seeing that such a hole can be used means that even if the intent wasn't there, things like 'this is a hole I opened to get milk or a newspaper, surely they wouldn't attack through that,' can still be used as attack channels.”

Security verification methods can be flexibly configured depending on the situation. Various options exist, such as repetitive automated verification, a trigger that activates whenever new code is deployed, or manual execution as needed.

This helps companies set optimal verification cycles tailored to their operating environment and security policies. CEO Park Se-jun explained, “We can do it repeatedly and continuously, or we can trigger it only when new code is deployed, or you can manually execute it if you feel like running it today.”

If even two out of ten patches are incorrectly applied, causing an outage, it leads to a major incident.

Therefore, defense is a highly difficult task that must protect citizens while preventing malicious attacks, and perfect correction and re-verification are essential when patching.

CEO Park Se-jun explained, “For defense, with patches, if we have ten, and two of them are patched incorrectly, causing an outage, that's a big accident. So, if ten patches are made, all ten must be perfect and safe.”

Credential stuffing is an attack technique that exploits the vulnerability of users reusing the same password or user ID across multiple services. If information is leaked from one place, unauthorized access to other services becomes possible.

Personalized phishing attacks based on leaked personal data are even more sophisticated and threatening. If sensitive information, such as hospital surgery and medical records, is exploited, severe secondary damage can occur.

Reusing passwords like a public commodity can lead to greater financial damage, such as bank account hijacking, so extreme caution is required. CEO Park Se-jun explained, “It’s a credential stuffing attack technique where people generally use the same password multiple times or have the same user ID. So, if one is compromised, they can try it elsewhere and access other services without even hacking them.”

In the early stages of AI adoption, attackers will always be one step ahead of defenders, leading to an expected rapid increase in AI-driven attack damage in the short term. It will be difficult for all companies to equip themselves with security systems in line with AI trends, allowing attackers to exploit opportunities and inflict significant damage.

The lack of verification for AI-generated code and AI learning from existing vulnerable codes, thereby perpetuating incorrect patterns, also serves as a major source of security vulnerabilities. CEO Park Se-jun said, “In the short term, not everyone can keep up with these trends, and attackers are always one step faster, so they will inflict a lot of damage when they have the chance. So, in the short term, I think the spread of AI-driven attacks and the scale of damage will rapidly increase.”

CEO Park Se-jun aims to prove the global competitiveness of Korean security companies through his vision of making the world safer with technology.