For years, hacking has required a certain level of technical expertise. Finding a weakness in a computer system is one thing but knowing how to exploit it is another. Now, artificial intelligence (AI) is beginning to narrow that gap. A person with limited coding knowledge can use AI to find vulnerabilities in software and work out how to take advantage of them. This is making cybersecurity companies nervous.The concern comes amid two developments this week. First, on October 24, Chinese startup Moonshot AI reduced the cyber-offensive capabilities of its Kimi K3.1 Pro model before releasing its weights publicly, likely on November 4. The company said it had removed cyber-related material from its training data and testing environments. In its internal tests, the company said the modified model scored 81 per cent on a benchmark measuring its ability to find and exploit vulnerabilities. The unmodified version scored 90.5 per cent.Later on October 29, French AI startup Mistral released Large 4, a model designed to help organisations find security weaknesses in their software. It is intended to help cybersecurity teams identify problems before attackers do. But the same capabilities could also be used by people looking for ways into computer systems.These two announcements reveal a problem that AI developers are increasingly having to confront: a model that can help a company protect its network may also help someone attack it.Cyberattacks are not new. What is changing now is the level of expertise needed to carry them.An experienced hacker can use AI to speed up work that might have taken them hours to finish. Someone with less experience can ask a model to explain some code or help identify a weakness. The model can even suggest ways to exploit it. While humans still need to be involved, but the barrier is coming down.There have also been indications of this shift. In June, Anthropic published an analysis of 832 accounts it had banned for malicious cyber activity. The report said that the proportion of actors classified as medium or high risk increased from 33 per cent to 56 per cent in a year.This does not mean AI is independently running cyberattacks. People are still directing these operations. But they now have a tool that can do some of the more complicated work for them.The concern about AI’s involvement in cyber threats also increases when models are released for public download. Users can run these models on their own computers without going through the company that developed them. This makes them useful to organisations that want greater control over their technology. It also makes it difficult for developers to monitor their use.Mistral’s approach illustrates the dilemma. Once a model is publicly available, the company cannot easily prevent someone from using it for malicious purposes. At the same time, restricting access could deny cybersecurity researchers a valuable tool.Companies need AI to test their defences. They cannot afford to wait for criminals to discover weaknesses first. Yet releasing increasingly capable models creates risks that developers may have little power to contain.One cannot find an easy answer to this dilemma. Making AI less capable at cybersecurity could weaken the people trying to defend computer systems. Releasing everything without restrictions could give attackers an advantage.For now, the stakeholders are not concerned that AI will suddenly decide to hack the internet by itself. But the concern is that people who once lacked the skills to carry out sophisticated attacks may find it easier to do so. And as AI models improve, that gap could shrink further.
Click Here For The Original Source.
——————————————————–
