OpenAI wants to submit new GPT versions to independent security experts

OpenAI wants independent experts to try to exploit unreleased versions of its GPT platform and services to make them more secure. The company said this to the British House of Lords.

OpenAI chimes in the response to the British House of Lords Where The Guardian writes about it that it would like rules to be introduced to have AI systems checked before release by independent experts in red teams. These are teams of security experts who try to crack systems, as malicious parties would do after release. “We believe it is important that our safety rules are externally validated by independent experts and that our decisions are at least partly based on independent safety and risk assessments, so that we can take appropriate measures
can develop.”

Red teaming currently only happens internally, but in many countries governments are considering how they want to regulate AI systems. OpenAI does not directly answer the question of how great the company estimates the risk that AI could be a threat to humanity. “As the capabilities of these models become more sophisticated, so too does the scope and severity of the risks they can pose, especially if they are directed by a malicious actor or if the model is not well aligned with human values. Strictly measuring Advances in potentially hazardous capabilities are essential to effectively assess and manage risk. We address this by researching and developing assessments for potentially hazardous capabilities, which range from simple, scalable, and automated tools to customized, intensive assessments that are carried out by human experts.”

OpenAI also says in the document that training large language models requires the use of copyrighted material. “Limiting training data to public domain books and drawings created more than a century ago may make for an interesting experiment, but would not produce AI systems that meet the needs of today’s citizens.”