GPT-5 Hacked in 24 Hours: Why the New OpenAI Model is Dangerous for Users

The new artificial intelligence model GPT-5 was 'hacked' just 24 hours after its release. This happened after its predecessor Grok-4 was also breached. NeuralTrust and SPLX (formerly known as SplxAI) reported that the raw version of GPT-5 is practically unusable due to serious security issues. This was reported by the news resource SecurityWeek.

Security Flaws in the GPT-5 Model

NeuralTrust used its EchoChamber system and narrative method to manipulate the model for malicious purposes without triggering protective mechanisms. This approach allowed the model to generate instructions for creating a Molotov cocktail.

'The attack successfully helped the new model create a step-by-step guide for making a Molotov cocktail,' the company says.

Furthermore, a group of researchers from SPLX also conducted their tests on GPT-5 using obfuscation attacks, specifically the StringJoin method, which allows masking a query as an encryption task.

NeuralTrust and SPLX found serious security flaws in the artificial intelligence model GPT-5, which could potentially lead to dangerous consequences in the future.