Archive — US — techcrunch
OpenAI’s latest AI models have a new safeguard to prevent biorisks
OpenAI has launched a new safety system to monitor its AI models, o3 and o4-mini, to prevent them from providing harmful advice on biological and chemical threats. The system is designed to refuse requests related to such risks and has shown to block 98.7% of risky prompts. However, concerns remain regarding OpenAI's overall safety measures and testing practices.
Key points
- OpenAI deployed a monitoring system to prevent its models from providing harmful advice related to biological and chemical risks.
- The system enhances safety measures in o3 and o4-mini, which show increased capability compared to previous models.
- Internal testing indicated a 98.7% success rate in blocking risky prompts.
- OpenAI acknowledges the need for human oversight beyond the automated monitoring.
- Concerns exist from outside researchers about the adequacy of safety priorities and testing time.
Why it matters
- As of 2025, AI safety continues to be a significant concern due to rapid advancements and potential misuse.
- Previous incidents involving AI-generated misinformation highlight the need for robust safety measures in AI development.
- The trend towards implementing automated safety systems in AI models is increasing in response to public and regulatory scrutiny.
OpenAI's proactive approach with its new monitoring system reflects a growing priority on AI safety, yet ongoing criticisms indicate that the balance between innovation and security remains a challenge in AI development.