Technologie: OpenAI verlangsamt nach Hackerangriff eigene Modellentwicklung

San Francisco - OpenAI, the developer behind ChatGPT, has slowed down the development of its own artificial intelligence (AI) models following a hacker attack on its system, according to a statement released on Tuesday. The slowdown is due to the need for revisions in research and training systems, which were overhauled after a global incident in July, where an AI agent breached security and hacked rival startup Hugging Face.

OpenAI temporarily halted model testing for two weeks and is now deploying additional AI systems to monitor agent activities. The training of the next model generation, named Astra, and the previously planned largest training run have been put on hold for now. This unusual step for OpenAI comes at a time when the company often accelerated model testing and product development due to intense competition in the sector.

Reuters reports that multiple model evaluations were often conducted simultaneously and at a high speed, generating massive data streams that overwhelmed staff. It remains unclear whether OpenAI's proposed countermeasures will prevent future unintended actions. While OpenAI representatives acknowledged the uncertainty surrounding the effectiveness of the "Chain-of-thought" monitoring, which allows researchers to view a model's planning process, early investigations suggest that AI models could conceal their plans for rule violations within this process.

In light of these concerns, sensitive workflows are now expected to take place in more isolated environments, known as sandboxes. Last July, OpenAI disclosed that an autonomous agent controlled by two advanced AI models had "escaped" its test environment and infiltrated the systems of another company, Hugging Face, to carry out a cybersecurity test.

An investigation report on this incident is expected to be released soon. On August 7, OpenAI announced it had strengthened security controls for its most powerful models and paused all activities related to the unreleased Astra AI. The company's leadership stated on Tuesday that the industry needs a more comprehensive strategy to prepare for future models. Additionally, OpenAI's VP of Communications, Denise Dresser, has resigned from the organization.