AI Misuse and the Future of Crime Prevention
AI as an accelerator for criminal innovation
For Kespite, the true turning point lies in the fact that AI has become accessible to everyone. A few years ago, advanced models were limited to specialists. Today, anyone with an internet connection can use powerful tools to generate images, text, audio or even entire identities. Open-source models can be downloaded and run locally, away from any oversight or monitoring.
He gave a simple contrast. When ChatGPT was launched in November 2022, many people saw it as a way to summarise documents or draft emails. Criminals, by contrast, immediately saw its potential for writing convincing phishing messages in any language, erasing the tell-tale spelling mistakes and awkward phrasing that used to make scams easy to spot. The same logic applies to image generation: while most users explore creative or playful uses, offenders exploit the technology to create synthetic child sexual abuse material where original content is harder to obtain.
Criminals are also extremely skilled at repurposing seemingly harmless techniques. Face-swapping, for instance, can be used to create sophisticated video calls in which an employee genuinely believes they are speaking to their CEO and is pressured into approving a multimillion-dollar transfer. “You might think the person who authorises this is foolish,” Kespite remarked, “but with urgency, stress and a convincing fake, many people can be manipulated.”
Deepfakes, cloned voices and the erosion of trust
One of the most troubling aspects of AI misuse is the way it undermines trust in audio and visual evidence. “Today you can no longer trust what you hear,” Kespite stressed. With just a few seconds of recorded speech, anyone can now clone a voice, type whatever text they want and generate a highly realistic audio message offline, without leaving any online trace.
He described a real-life example where a school principal in the United States was temporarily suspended and threatened after fake audio recordings of them making offensive remarks circulated online. Only months later was it confirmed that the clips were deepfakes. By then, the reputational damage had already been done. The same techniques are increasingly used to simulate kidnapping calls or emergency situations, with scammers cloning the voice of a family member and calling relatives to demand urgent payments.
The visual dimension is just as disturbing. AI can now undress individuals in photos, producing fake nudes from perfectly innocent images. Teenagers and young adults, particularly girls, are targeted in so-called revenge porn scenarios, where ex-partners manipulate images to humiliate and blackmail them. AI can also turn manga-style characters into photo-realistic young women, blurring even further the distinction between real and synthetic victims.
For investigators, this creates a new nightmare. Kespite described a scenario in which law enforcement officers discover a trove of apparent abuse images and invest time and resources trying to identify and rescue the victims. In reality, some or all of those “victims” may not exist at all. “You think you are saving someone, but that person doesn’t exist,” he warned. The opportunity cost is huge: while teams chase phantom victims generated by algorithms, real children may remain unidentified and at risk.
Large language models: from assistants to criminal agents
The misuse of large language models (LLMs) is another major concern. Commercial systems include safeguards and content filters, but Kespite pointed out that criminals are creative in bypassing them. They reframe malicious intent as fiction, research or creative writing, or simply abandon controlled platforms in favour of open-source models they can modify themselves.
He mentioned a real case in which a suspect involved in an attempted bombing had asked an LLM about explosives and triggering mechanisms. Even if the model did not provide full instructions, the incident shows how natural it has become for offenders to turn first to AI for guidance. And when guardrails do block them, they can simply set up local copies of models, strip out the safety layers and create crime-friendly GPTs.
The real shift, however, lies in automation. By combining LLMs with tools for planning and execution, criminals can build agents that do more than answer questions. They can devise step-by-step strategies, create cryptocurrency wallets, split funds, use mixing services to anonymise transactions, or scan the internet for vulnerabilities and launch cyberattacks. A single criminal can orchestrate an entire “army” of AI agents probing systems, exploiting weaknesses and laundering money with minimal human input.
“You give the model tools, you ask for a plan, and it executes,” Kespite summarised. In such a context, the traditional image of a lone hacker typing furiously at a keyboard feels increasingly outdated. The real threat is a semi-autonomous ecosystem in which AI handles most of the operational legwork.
Keeping up: speed, collaboration and a new mindset
Despite the alarming picture, Kespite’s conclusion was not fatalistic. He argued that the response rests on three pillars: technology, partnership and regulation. Law enforcement agencies and public authorities have to adopt advanced tools themselves, to avoid falling permanently behind. Partnerships and collaboration with academia and the private sector is essential, because no single actor can be an expert in every field, from deepfake detection to cryptography and online platforms.
Regulation, in his view, must both constrain harmful uses and empower investigators to use AI lawfully and effectively. But above all, he insisted on a change of pace and culture. Traditional projects that take three years to deliver are no longer fit for purpose. “If you start an AI project today that delivers in three years, it will already be obsolete,” he warned. With criminals, it’s an arms race.
He closed by tackling one of the most persistent misconceptions. “AI does not replace humans. What replaces people are people who use AI,” he said. In the age of artificial intelligence, the real divide is not between machines and humans, but between those who adapt and those who allow criminals to stay several steps ahead.
Image credit : Zulfugar Karimov - Unsplash
Image credit : Cleverdis