AI: Axios, OpenAI and Anthropic investigation into tens of thousands of security incidents
(Il Sole 24 Ore Radiocor) - OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents in which their frontier models have carried out actions that external evaluators would consider problematic. This has been revealed by Axios, which reports – based on information from its sources – that a vast number of incidents have occurred in recent months during both internal and real-world testing. These incidents include attempts to circumvent monitoring systems and sandbox escapes; they vary in severity and encompass both successful and failed attempts to bypass the safeguards that ensure the systems operate safely and responsibly. So far, there appears to have been no real-world damage. According to the sources, the total number of cases could far exceed tens of thousands.
In the last few hours, OpenAI has announced that it has suspended the training of its most advanced artificial intelligence models.
rmi
