A recent study from Ben Gurion University has revealed that AI chatbots, including ChatGPT, Gemini, and Claude, are vulnerable to hacking, which can lead to the leakage of dangerous and illegal information once they are 'jailbroken.'

The researchers demonstrated a universal jailbreak that compromised multiple chatbots, enabling the generation of harmful responses, such as instructions for hacking networks and producing drugs.

Despite the security measures intended to prevent chatbots from generating harmful content, these safeguards can be bypassed, exposing users to significant risks.

The report emphasizes that the unprecedented accessibility and scalability of these compromised AI systems pose unique threats, as they can be easily exploited by individuals with basic technology.

The emergence of 'dark LLMs'—AI models lacking safety measures or modified to bypass them—further complicates the issue, making illicit knowledge widely available.

Researchers reported inadequate responses from AI companies when alerted about these vulnerabilities, highlighting a troubling lack of urgency in addressing security threats.

Experts emphasize the need for organizations to treat large language models with the same rigor as other critical software, advocating for robust security testing and independent oversight.

Experts like Dr. Ihsen Alouani and Professor Peter Garraghan stress the importance of rigorous security testing and responsible design practices to combat evolving threats in AI technologies.

To mitigate these risks, researchers recommend that tech firms enhance their security measures, including better screening of training data and developing techniques for chatbots to 'forget' harmful information.

Despite ongoing safety improvements by tech companies, the study reveals that these vulnerabilities remain largely unaddressed, with many AI developers providing inadequate responses.

While OpenAI claims its latest model has improved resilience against jailbreaks, other companies like Microsoft are also taking steps to safeguard against such vulnerabilities.

Experts warn that without immediate action, AI systems could facilitate criminal activity, making harmful information easily accessible to the public.