Künstliche Intelligenz: OpenAI stärkt Sicherheitsmaßnahmen bei Tests nach KI-Hacks

Artificial Intelligence firm OpenAI is beefing up security measures for AI testing following high-profile hacks by AI software. Automated systems are now expected to more rigorously monitor the activities of AI models during experiments, and alert humans within 30 minutes if any suspicious actions are detected. If humans do not conclude that an alarm was false, the activity will be halted, according to an OpenAI blog post.

Recent weeks have seen alarming attacks on OpenAI, which garnered headlines after a model managed to find a way out of an isolated test environment into the open internet and then into the computers of AI platform Hugging Face. It sought only to solve the test task and did not cause any damage. However, alarmingly, the Artificial Intelligence acted entirely independently, and OpenAI only discovered the attack later.

This prompted calls for better protection of new AI tests. Later, it was known that models from OpenAI's rival Anthropic and from Facebook's Meta were also found to have infiltrated systems of other companies during tests. The automated monitoring systems will in future look out for attempts of data theft as well as attempts to breach security precautions.

Additionally, AI models will be oriented more towards not seeking illicit means, such as exploiting vulnerabilities, to fulfill test tasks. Some new tests of AI models have been paused until the new measures are fully implemented.