Opeli claims that he has implemented a novel monitoring system for the latest AI, O3 and O4-Mini reasoning models in order to obtain hints related to biological and chemical threats. The system aims to prevent models from advice that could instrument someone about conducting potentially harmful attacks, According to the Openai security report.
The company says that O3 and O4-Mini are a significant augment in possibilities compared to previous OPENAI models, and thus constitute a novel risk in the hands of bad actors. According to OPENAI internal references, O3 is more qualified in answering questions about the creation of some types of biological threats. For this reason, and reduce other risks-openai has created a novel monitoring system, which the company describes as “a monitor of security focused on security.”
The monitor, used to order for reason regarding the principles of OpenAI content, works on O3 and O4-Mini. It has been designed to identify hints related to biological and chemical risk and to instruct models to refuse to give advice on these topics.
To establish a base line, OpenAi told red teams to spend about 1000 hours, marking “dangerous” conversations related to the registration of O3 and O4-Mini. During the test, in which Opennai simulated the “blocking logic” of its safety monitor, the models refused to respond to risky hints 98.7% of time, according to OpenAI.
Opeli admits that his test has not taken into account people who can try novel hints after blocking by the monitor, so the company claims that it will continue to consist in monitoring people.
According to O3 and O4-Mini, they do not exceed the threshold of “high risk” OpenAI for Zabis. However, compared to O1 and GPT-4, OpenAI claims that the early versions of O3 and O4-Mini have proved to be more helpful in answering questions about the development of biological weapons.
The company actively follows how its models can facilitate malicious users of the development of chemical and biological threats, in accordance with the recently updated OPENAI Readiness framework.
Opeli is increasingly involved in automated systems to reduce the risk of its models. For example, to prevent Native GPT-4O image generator from the creation of children’s sexual materials (CSAM)Opeli claims that it uses a reasoning monitor similar to the one that the company has implemented for O3 and O4-Mini.
However, several researchers aroused OpenAI’s fears not priority treated as much as it should. One of the company’s red partners, a meter, said that he has relatively little time to test O3 with reference for duplicitous behavior. Meanwhile, Opeli decided not to issue a security report on its GPT-4.1 model, which was published at the beginning of this week.