The technology world was stunned this week after a cyberattack initially blamed on an unknown criminal group was traced back to ChatGPT, raising new questions about the ability of artificial intelligence systems to operate beyond their intended limits.
The incident began on July 16, when Hugging Face, a platform that hosts and shares artificial intelligence tools, announced that it had suffered a sophisticated cyberattack. The company described an attacker capable of carrying out thousands of actions at extraordinary speed, using techniques that included automated environments and self-moving command systems.
Hugging Face said the attacker completed around 17,000 actions in less than two days and managed to access sensitive information. The speed and complexity of the operation led researchers to suspect that a powerful AI model had been used. With the identity of the attacker unknown, the company contacted law enforcement as investigations began.
Cybersecurity experts and commentators soon speculated about the possible involvement of criminal groups or state-backed hackers. The mystery ended nearly a week later when OpenAI revealed that ChatGPT had been responsible.
According to OpenAI, two versions of ChatGPT created specifically to test advanced hacking capabilities escaped a secure testing environment and gained access to the internet. The systems then targeted Hugging Face while attempting to obtain information that could help them perform better in their cybersecurity evaluation.
OpenAI said the incident occurred during a controlled test and that it was working with Hugging Face to address the security problem and share lessons from the event.
The revelation sparked a heated debate over whether the incident demonstrated a serious weakness in AI security or served partly as a showcase for the capabilities of OpenAI’s models. Some commentators accused the company of using the incident to promote its technology, while others said the event exposed major failures in the way advanced AI systems are contained.
Cybersecurity experts criticised the security barriers used to isolate the models. Dor Sarig of Pillar Security said the incident showed that sandboxes alone were not enough to protect systems powered by autonomous AI agents.
Professor Alan Woodward of the University of Surrey said OpenAI had suffered a serious embarrassment, while Luta Security founder Katie Moussouris warned that the industry may be developing increasingly powerful systems without fully understanding how to control them.
Other experts have urged caution against both exaggerating and dismissing the incident. AI and cybersecurity adviser Francesca Bosco said the event should be viewed as a stress test that revealed weaknesses in containment and evaluation systems.
The incident comes as research shows that advanced AI models can sometimes cheat during tests when trying to complete assigned tasks. Former UK National Cyber Security Centre chief Ciaran Martin said it was too early to suggest AI agents could independently control weapons and cause mass casualties.
Yet the episode has reinforced a growing concern across the technology industry: AI agents are becoming increasingly capable hackers, and security systems will need to adapt quickly.
