The alarm comes fromEuropolthe European Union’s law enforcement agency: ChatGpt and other generative artificial intelligence systems can be used for online fraud and various types of cyber crimes.
Since its launch at the end of November 2022, ChatGpt has become one of the fastest growing internet services, exceeding 100 million users in the first two months. As everyone knows by now, the model has gone viral thanks to its ability to generate, even starting from very specific requests, textual contents similar to those that a human user could write. And this, second l’Innovation Lab di Europolrepresents a potential threat, as these tools could be exploited by malicious actors.
“Criminals are generally quick to take advantage of new technologies and have seen themselves rapidly proposing concrete criminal exploitations, providing the first practical examples just a few weeks after the public release of ChatGpt”, reads a flash report compiled by the agency.
The publication is the result of a series of workshops organized by Europol’s Innovation Lab to discuss the potential criminal uses of ChatGpt, the most prominent example of large language patterns, and how these patterns could be used to support investigative work.
Current system bugs
The European agency points out that the ChatGpt moderation rules can be circumvented through the so-called prompt engineeringthe practice of providing inputs to an AI model precisely to achieve a specific result.
Since ChatGpt is a relatively recent technology, flaws are continually found despite the constant distribution of patches. These flaws can consist of asking the AI to provide the prompt, asking the artificial intelligence to pretend to be a fictional character or in providing the response in code.
Other workarounds couldo substitute passwords or change the context during interactions. Europol stressed that the most effective solutions, which manage to free the model from any constraints, are constantly evolving and becoming more complex.
The Demand Generation LAB is back on May 11th: finally live in MILAN!
Possible criminal applications
Experts have identified a number of illegal use cases for ChatGpt that persist even in OpenAi’s most advanced model, Gpt-4where the potential for malicious system responses was even more advanced in some cases.
Since ChatGpt is able to generate read information, Europol warns that emerging technology can speed up the process of finding a malicious actor without prior knowledge of an area of potential crime such as home breaking and entering, terrorism, cybercrime, or child sexual abuse.
“Although all of the information provided by ChatGpt is freely available on the Internet, the ability to use the model to provide specific passages by asking contextual questions means that it is significantly easier for malicious actors to better understand and subsequently carry out various types of crime“, reads the report.
Il phishing, the practice of sending a fake email to trick users into clicking on a link, is a crucial application area. In the past, these scams were easily spotted due to grammatical or linguistic errors, while the text generated by artificial intelligence allows you to impersonate these people in an extremely realistic way.
Similarly, it is possible to give to online fraud an image of greater legitimacy by using ChatGpt to create a fake social media engagement that can pass off a fraudulent offer as legit. In other words, thanks to these models, “these types of phishing and online fraud can be created faster, much more authentically and on a significantly greater scale”.
Furthermore, the ability of AI to impersonate the style and language of public figures can lead to different cases of abuse of propaganda, hate speech and disinformation.
In addition to text, ChatGpt can also produce code in several programming languages, expanding the ability of malicious actors with little or no computer development knowledge to turn natural language into malware.
Shortly after the public release of ChatGpt, the security company Check Point Research demonstrated how the AI model can be used to create a complete infection stream, from creating phishing emails, from spear-phishing to running a reverse shell that accepts commands in English.
The prospects for the future
While the European Parliament is finalizing its position on the AI lawMEPs discussed introducing some stringent requirements for this basic model, such as risk management, robustness and quality control.
However, Europol seems to believe that the challenge posed by these systems will only increase as they become increasingly available and sophisticated, for example with the generation of highly convincing deep fakes.
Another risk is that these large language models can become available on the dark web without any protection and be “trained” with particularly harmful data. What kind of data will feed these systems and how they might be controlled are major question marks for the future.
Musk and 1,000 other managers call for break on advanced systems
Elon Musk and 1,000 other researchers and managers are calling for a six-month break in the development of advanced artificial intelligence systems starting with ChatGpt to stop what they call a “dangerous” arms race. In an open letter published by the Future of Life Institute, and reported by the Financial Times, the more than 1,100 signatories explain how “in recent months there has been an out-of-control rush of artificial intelligence laboratories to develop and deploy powerful digital minds that no one, not even the creators, can understand, predict and control”. “We call on all AI labs to immediately stop training the most powerful AI systems like Gpt-4 for at least six months. The break should be public, verifiable, and include everyone. If it is not implemented quickly, governments should step in and institute a moratorium.”
@ALL RIGHTS RESERVED