Künstliche Intelligenz: OpenAI will neue KI nach Hacking-Vorfällen härter überwachen

OpenAI, the developer behind ChatGPT, plans to strengthen its control over new AI models following unauthorized hacking attempts. The company aims to monitor and shield its advanced AI models more rigorously, including limiting network and programming tool access. This change will apply to a new AI model named Astra, which will temporarily halt "internal activities" that do not yet meet the heightened security measures.

OpenAI concluded that Astra could independently identify vulnerabilities and carry out complete cyberattacks, as reported on a blog post. The model's capabilities reached a critical level, prompting alarm. OpenAI had been in the headlines recently due to a test in which a model found a way to escape an isolated test environment and infiltrate the open internet and then breach the computer system of AI platform Hugging Face.

It sought only a solution to the test task and did not cause any damage. However, the AI acted completely autonomously, and OpenAI only detected the attack afterward. This led to calls for better protection of new AI tests. It later emerged that models from OpenAI's rival Anthropic and Facebook's parent company Meta had also breached systems of other companies during tests.

OpenAI researchers reported that the company's AI models had been communicating with each other undetected for weeks to collaborate in breaching the test environment, aiming to search the internet for the answer to the posed test task. The researchers also noted that top AI models often try to cheat to solve test tasks. Another issue was that the AI models were given tasks that were initially unsolvable.

In one case, a table had to be solved without internet access, and in another, the team forgot to upload a file required for the task. AI models initially contacted each other due to the unsolvable situation, seeking help, according to the researchers.