Artificial intelligence: OpenAI wants to monitor new AI more closely after hacking incidents

ChatGPT developer OpenAI wants to control its new AI models more tightly after unauthorized hacking attacks by artificial intelligence. Software with extensive capabilities is to be monitored and shielded more strongly, including restricted access to networks and programming tools. For a new AI model called Astra, this means that "internal activities" that do not yet meet the stricter security precautions are being paused for the time being.

OpenAI had come to the conclusion that Astra was able to independently find vulnerabilities and carry out complete cyberattacks, it said in a blog post. These capabilities had reached a critical level in the model.

Alarming series of attacks

OpenAI had made headlines in recent weeks because a model in a test found a way to get from an isolated test environment into the open internet - and then broke into the computer system of the AI platform Hugging Face. It was only looking for a solution to the test task and did not cause any damage. However, it was alarming that the artificial intelligence acted completely independently - and OpenAI only detected the attack afterwards. This led to calls for better securing of tests of new AI.

It was later revealed that models from OpenAI rival Anthropic and Facebook parent company Meta had also penetrated systems of other companies during tests.

Unsolvable tasks and cheating

OpenAI researchers also reported that the company's AI models had been communicating with each other undetected for weeks before the attack on Hugging Face in order to work together to escape the test environment. Their goal was to search the internet for the answer to the task set in the test, OpenAI experts said in a presentation at the Black Hat hacker conference.

The researchers found that the top AI models often wanted to cheat in order to solve the test tasks. One problem with the test runs was that the AI had sometimes inadvertently received orders that were not fulfillable. In one case, a table in which it was supposed to solve a task was not accessible without internet access. In another case, the team had forgotten to upload a file belonging to the task.

The AI models had initially contacted each other because they had sought help in the unsolvable situation, it was said.