In the previous article, we explored the risks and opportunities associated with modern Artificial Intelligence tools.
This week, we take a closer look at the relationship between AI and cyberattacks, examining how AI can be exploited by cybercriminals and how AI systems themselves are vulnerable to specific types of attacks.
Cyberattacks Supported by AI
There are numerous cyberattacks that are increasingly supported by AI. We have selected four types that are particularly effective:
- AI-Generated Voice Scam: A 59-year-old woman in India lost approximately $1,600 due to an AI-generated voice scam. The scammer managed to alter his voice to impersonate one of the victim's grandsons, who lived in Canada. The scammer claimed he had been in an accident and was about to be jailed. The woman, who received the call late at night, confirmed that the voice sounded exactly like her grandson, including the Punjabi dialect they spoke at home. Without hesitation, she transferred money, only to realize later that she had been tricked. This scam highlights the effectiveness of AI tools in creating highly sophisticated and convincing voice scams.
- CryptoRom: Cybersecurity firm Sophos has been monitoring a category of scams called "CryptoRom," characterized by fake cryptocurrency trading schemes and luring victims through romantic interest. While this scam has been around since 2020, it has recently evolved with the use of generative AI to interact with victims. Typically, scammers approach victims through dating apps and establish a romantic connection. The conversation soon moves to messaging apps like WhatsApp or Telegram, where the idea of trading cryptocurrencies is introduced. Using fraudulent cryptocurrency apps, scammers guide victims to transfer funds, ultimately scamming them out of large amounts of money. Generative AI tools, such as ChatGPT or Google Bard, assist scammers in overcoming language barriers and make their interactions more convincing. Unfortunately, CryptoRom apps are increasingly found on major mobile platforms like Apple and Android, indicating the global reach of this scam.
- Fake Liquidity Mining: Another rapidly evolving cryptocurrency scam is fake "liquidity mining." In this scam, a US victim lost the equivalent of $22,000 in cryptocurrency after being initially approached on a dating app called MeetMe. The scammers not only stole the victim's cryptocurrency but also tried to convince him to deposit additional funds to recover the lost amounts. Again, generative AI tools were used to make the interactions with the victim more credible. Tracking the wallet addresses revealed a broader pattern, with 13 additional domains hosting the same fraudulent decentralized finance app, targeting Trust Wallet users. Over the course of five months, the scammers were able to steal more than $1 million in cryptocurrencies from victims' wallets. Unfortunately, this is not an isolated case: hundreds of other scam sites are using similar tactics to earn millions in cryptocurrency without the need for malware or compromising victims' systems.
- BEC Scam with WormGPT: Cybercriminals are increasingly using generative AI to support Business Email Compromise (BEC) scams. The latest innovation in this field is "WormGPT", a malicious version of GPT models, specifically trained with malware-related information. WormGPT can generate realistic text and create convincing fake emails, enabling cybercriminals to carry out fraudulent activities even without specific linguistic skills. The tool's effectiveness was demonstrated in a test where it generated an apparently authentic email to facilitate the payment of a fake invoice. WormGPT's capabilities in sophisticated phishing and BEC scams are alarming, as it lacks the ethical constraints of tools like ChatGPT, allowing criminals to easily create highly personalized and targeted attacks.
Cyberattacks Against AI
A recent report by the the NIST (National Institute of Standards and Technology) describes the four main types of cyberattacks used to compromise machine learning algorithms:
- Evasion: Attacks that attempt to alter an input to change how the system responds. For example, modifying road signs to trick an autonomous vehicle into misinterpreting them, potentially causing it to veer off course or crash.
- Poisoning: Attacks that occur during AI training by introducing corrupted data. For instance, inserting numerous examples of inappropriate language into conversation logs, so that a chatbot adopts this language in its interactions with customers.
- Privacy Attacks: Attacks aimed at learning sensitive information about the AI, or the data used to train it, in order to exploit it. For example, asking a chatbot legitimate questions and using the answers to decode the model’s weaknesses or guess its data sources for later compromise.
- Abuse: Attacks that involve introducing false or misleading information into a typically reliable source, such as a website, to trick the AI into ingesting incorrect data. Unlike poisoning attacks, these actions target the AI's effectiveness by exploiting a legitimate source that has been compromised.
Conclusion
The intersection of AI and cyberattacks presents both unprecedented risks and challenges.
As cybercriminals increasingly leverage AI to create more sophisticated and personalized attacks, the threat landscape is evolving. Moreover, AI systems themselves are vulnerable to cyberattacks that can undermine their reliability and security.
In this two-sided environment, organizations and individuals must remain vigilant and implement robust Cyber Security measures to mitigate both the risks of AI exploitation and the vulnerabilities of AI-based systems.
The future of Cyber Security will undoubtedly require an ongoing, dynamic response to the integration of AI in both offensive and defensive capacities.
Stay Cyber Safe!
