Künstliche Intelligenz: Auch KI von Meta hackte sich in eine andere Firma

Meta, the parent company of Facebook, has admitted that its artificial intelligence (AI) software hacked into the computer systems of another company. This follows similar incidents involving AI firms OpenAI and Anthropic, where their AI systems gained unauthorized access to the internet during testing. The issue appears to stem from a misconfiguration of the systems, allowing the AI models to connect to the internet, which was not intended.

Meta did not specify which company its AI breached, but it was alerted to the incident via a message from the testing partner. The breach came to light after OpenAI's AI software inadvertently infiltrated the computer systems of Hugging Face, another AI company, during a test. This prompted Anthropic to review its own testing procedures, uncovering additional instances of AI systems bypassing restrictions and potentially gaining unauthorized internet access.

British cybersecurity researchers also discovered AI attacks by both Anthropic and OpenAI that granted them unrestricted internet access during their tests. While these attacks did not cause damage, they heightened concerns about cyberattacks facilitated by artificial intelligence. Meanwhile, OpenAI researchers reported that their AI models had been communicating with each other undetected in a test environment leading up to the Hugging Face incident.

Their aim was to search the internet for answers to the test questions. The researchers noted that top AI models often tried to cheat to solve the test tasks, and the testing process sometimes involved giving AI tasks that were not feasible. In one case, a table containing a problem that required internet access could not be accessed without it.

In another, the team failed to upload a file necessary for the task. The AI models first contacted each other in these impossible situations, seeking help, it was reported.