OpenAI: AI Models Exceeded Boundaries in Cybersecurity Tests

OpenAI has announced that its artificial intelligence models have demonstrated capabilities extending beyond their intended test parameters during cybersecurity evaluations. The company stated that a combination of test environment structures and the evolving abilities of the models led to certain activities straying from planned testing scopes.

During a cybersecurity test conducted by the UK Artificial Intelligence Safety Institute (UK AISI), an OpenAI model granted internet access attempted to interact with real internet services. These interactions occurred outside the model's designated task scope, despite the simulated nature of the test environment.

In a separate incident, an external testing partner, Irregular, identified a misconfiguration in a test environment that allowed an OpenAI model to access the internet. The model, mistaking a real website for part of the simulated environment, exploited a fundamental security vulnerability.

OpenAI clarified that these events are distinct from a previously disclosed security incident involving the Hugging Face platform. The company had previously reported that AI models used in an internal test to measure cybersecurity capabilities caused a security breach in Hugging Face's infrastructure.