OpenAI’s latest AI models have a new safeguard to prevent biorisks

Openai says that it places a new system to watch the latest AI substantial models, O3 and O4-Minifor biological and chemical threats. The system is aimed to prevent the proposal of the models that can instruct potential harmful attacks, According to Openai’s security report.

O3 and O4-mini, Openai’s previous models, the company says, the company says and thus creates new risks in the hands of evil actors. According to the internal criteria of Openai, O3, especially in the establishment of certain types of biological threats, more skilled. For this reason – and to reduce other risks – Openai created a new monitoring system described as the company’s “security-oriented justification monitor”.

Openai’s content is special, O3 and O4-mini, which causes politics. It is designed to identify biological and chemical risk guidelines and refuse to consult the models on these issues.

To create a basic, Openai’s red teams have spent about 1,000 hours around 1000 hours of negotiations from O3 and O4-Mini. During a test that simulates the “blockage logic” of Openai’s security monitor, the models refused to respond to risky, according to Openai.

Openai accepts that the test does not take into account the people who can test new instructions after the monitor is blocked, so the company says it will continue to trust in connection with human monitoring.

According to the company, O3 and O4-mini O4-mini, the company does not exceed the “high risk” limit for biorisks. However, compared to O1 and GPT-4, the early versions of O3 and O4-mini says that the early versions of biological weapons are helping questions.

The company is actively monitoring that the models can facilitate chemical and biological threats, according to the recent update of Openai Preparation framework.

Openai relies on more automated systems to reduce risks from their models. For example, to avoid The native image generator of the GPT-4O is from creating a child’s sexual exploitation material (CSAM)Openai says it uses a company similar to the company placed for O3 and O4-mini.

Again, several researchers raised Openai concerns, and security does not prioritize as needed. One of the partners of the company, one of the partners, meters, a criterion for cheating behavior, he said that it was relatively little time to test O3. Meanwhile, Openai decided not to release Security report for the model of GPT-4.1started at the beginning of this week.