Is it time to start worrying about AI hackers?

AI attacks will first determine the target, then obtain feedback, and continuously optimize and improve. If one day AI can develop completely novel solutions or even begin to optimize solutions, its lethality will be huge.

System ambiguity is the current barrier against AI hackers

For games like Go, the rules, goals, and feedback are all simple and certain. , there is nothing external to disrupt the situation, losing or winning is feedback. GPT-3 AI can write coherent articles because the world it lives in is only text. Because of this, many current hacker reward activities are conducted in simulated environments, all artificial, all constrained, and all rules are prepared for AI.

The fuzziness of the system is the key. We can input all global tax laws into AI, because there are formulas in tax laws to determine how much tax is owed. Even so, all laws have ambiguities. Blur is difficult to translate into code and tricky for AI to handle. So even with AI, there will still be a need for a lot of tax lawyers in the foreseeable future.

Most human systems are more ambiguous. Could AI be used to hack real-world sports games (such as hockey games)? I'm afraid it's difficult. The AI ​​must understand the rules of the game, but that's not enough. It also needs to understand human physiology, the aerodynamics of clubs and balls, and more. Of course we can't say it's absolutely impossible, but it's definitely much harder than dealing with Go.

In the short term, it is precisely because social systems are extremely complex and ambiguous that the system’s defense against AI hackers is enhanced. It is unlikely that AI sports hackers will appear in the future. If robots participate in games in person, or one day in the future AI is so advanced that it can understand the vast world from all dimensions, then there may be AI sports hackers born.

Similarly, casino AI hackers and legal process AI hackers are not easy to develop. It will take a long time for AI to simulate the work of individuals or groups, and it is not possible for the time being.

But don’t underestimate the non-linear evolution capability of AI

The world is full of AI hackers. This is a science fiction problem, but it is not a stupid science fiction problem. AI is improving rapidly, but its progress is unstable and incoherent. Problems we thought were difficult may turn out to be easy, and problems we thought were easy may turn out to be difficult. A few decades ago we might have laughed at AI and thought it would be impossible for it to conquer Go because Go was too complex: not that the rules were complex, but that there were too many possible moves. What now? AI has already won Go.

So, we should think about the problem of AI hackers in advance and prepare plans early.

First of all, what we should pay attention to may be AI hackers in the financial system, because the rules of the financial system are easily exploited by algorithms. High frequency trading algorithms are a good example and will become more complex in the future. We can arm AI with real-time global financial information, add global laws and regulations, add news information, and add other relevant information, and then assign AI to complete a goal, pursuing "maximum legal profit" or "obtaining the largest possible profit". profit". At that time, there may be some hacking behaviors that are beyond human understanding, which means that humans will not even realize that hacking behaviors have occurred. Of course, such a thing seems a bit far away now.

In the short term, AI may cooperate with humans to complete hacking attacks. AI can find loopholes, and experienced accountants or tax lawyers can collaborate and use their rich experience to determine whether loopholes can be exploited for profit.

The lethality of mature AI hackers is immeasurable

Throughout human history, hacker attacks are generally human behaviors. If AI starts attacking, everything will change. AI will not be limited to certain methods and will not be subject to many restrictions like humans. AI doesn't rest at all, they think like aliens, and it can hack systems in ways humans can't predict.

Computers enhance attacks in four dimensions: speed, scale, scope and complexity. AI can increase attack power many times.

Speed: It may take months or even years for humans to launch an attack, but AI does not need to. It may take days, hours or even seconds. For example, what would happen if you fed the entire U.S. tax code into an AI and told it to find a plan to minimize someone's tax bill? If it is a multinational company, AI may be used to analyze global tax laws and reduce taxes for the company.

Scale: Once the AI ​​finds a vulnerability, it may exploit it at an extremely large scale. When AI begins to break into the financial system, they will dominate the entire industry. At that time, credit markets, tax laws, and general laws may all tilt in favor of the rich. AI will worsen inequality. The first AI financial hackers may not be developed by researchers pursuing equality, but by multinational banks, hedge funds, and management consultants.

Scope: Society does have a plan to deal with hackers, but it mainly deals with human hackers and attacks at human speed. If AI can quickly find hundreds of loopholes in tax laws, how can the country cope with it? The speed of amending tax laws simply cannot keep up.

Complexity: With the assistance of AI, human hackers can launch complex attacks whose complexity exceeds the level of unassisted human understanding. AI’s sophisticated statistical analysis can reveal relationships between variables and find hidden vulnerabilities that even the best strategists and experts may not be aware of.

When AI launches attacks at ultra-fast speed, ultra-large scale, ultra-wide scope, and ultra-high complexity, society may not be able to handle it at all. Will AI break through human limitations and even dominate society? Some AI researchers are indeed worried. Although it sounds like speculation now, early prevention is still necessary for humans. (Xiao Ge)

The above is the detailed content of Is it time to start worrying about AI hackers?. For more information, please follow other related articles on the PHP Chinese website!