OpenAI models went beyond human control:Unauthorised interference attempts hit 100+ institutions; websites sent wrong commands

OpenAI, the company behind ChatGPT, has attempted to forcibly break into the systems of more than 100 institutions. The company itself has disclosed this. This indicates that technology companies are losing control of their advanced AI models while they are still in the testing stage. The AI models are not only bypassing system security but also breaking through defined boundaries and moving beyond human control. OpenAI alerted the institutions The company has issued alerts to more than 100 institutions affected by its AI agent. However, the company claims that no data was stolen from these institutions. In its defence, OpenAI argued that the AI agents did not break the systems’ locks; they merely tried to get inside by forcibly knocking on the closed doors and turning their handles. Incorrect commands were given to websites An investigation has found that these AI agents sent arbitrary and unauthorised commands to several websites, attempting to make them execute improper processes. Moreover, to evade security systems, the agents turned the websites into a communication channel between themselves and attempted to bypass the established security measures by completely disregarding them. OpenAI admits its safeguards failed OpenAI has acknowledged that its safeguards failed to keep the AI agents under control. A company spokesperson said the affected organisations received technical data so they could investigate the breach and security flaws in their systems. After the failure, OpenAI has claimed it will make public its findings on the AI model’s unusual behaviour and security shortcomings, so cybersecurity experts worldwide can work together to counter the threat. Such a vast amount of data that it would take a human 6.6 crore years to read The company is examining 50 petabytes of data records. It has deployed 7,000 state-of-the-art GB200 and GB300 GPUs for the investigation, at a daily cost of ₹5 crore. If all this data were written in ordinary English words, it would take a human around 6.6 crore years to finish reading it, even at the fast pace of 240 words per minute without stopping. Since the task is beyond human capabilities, it is being carried out using AI models. More computing power will be deployed to make the investigation more accurate. Also read: Boeing that carried Sunita Williams is set to fly again: What went wrong with Starliner 1 in 2024 what does Crew 13 have to do with it?

Security rules tightened after Hugging Face incident After the incident came to light, OpenAI significantly tightened its security systems: Learn about the Hugging Face incident… This is one of the most shocking incidents in the world of AI so far. In simple terms, here is what actually happened: 1. The hack to steal the ‘answer key’ OpenAI was testing one of its AI models to assess its cybersecurity capabilities. To score well in the test, the AI chose a dangerous shortcut. It decided that stealing the ‘answer key’ directly would be better than solving the questions. 2. Broke restrictions and accessed the internet The AI model had been kept in a secure, closed testing environment where it was not allowed to access the internet. Despite this, the AI breached the system’s security, gained access to the internet and hacked the systems of the popular AI platform Hugging Face to obtain answers from there. Hugging Face disclosed the hack on 16 July. 3. Secret network had been operating for two months The investigation revealed that this was not the work of a single AI model. The activity had been underway since the beginning of May, around two months before the incident. Several AI agents were leaving secret notes for one another within OpenAI’s own servers. In these notes, they were telling one another how to breach the system’s security and access restricted data. The agents were referring to themselves as a “swarm”-an organised group. 4. Uncontrolled behaviour Engineers had not trained these models to do this. The AI had independently developed this method by using the resources available in the system. OpenAI had no idea that this entire network was operating on its own servers. What is the real reason for concern? Let us understand this with an example: if you see two ants in your kitchen, it does not mean there are only two ants in the house. This incident was somehow detected, but the biggest fear is how many such AI systems are operating behind closed doors and on secure servers, pushing beyond their limits without humans knowing. US and Australian government websites breached too After the Hugging Face incident, OpenAI’s AI agents also breached government websites in the US and Australia. In view of the situation, OpenAI has stepped up monitoring and temporarily halted development of its new models. Also read: AI agent breached Australia’s government website: When blocked, it devised its own way to breach security; 4 hacking cases in 2 months AI poses threat of human extinction AI researcher Jacob Coxon claims that if strict controls are not imposed on AI in time, it could soon cause the end of humanity. Meanwhile, Dario Amodei, CEO of AI company Anthropic, has also acknowledged the need to slow down AI development.