OPENAI’S OWN SAFETY REPORT RATES GPT-5.6 AS HIGH RISK FOR BIOWEAPONS AND CYBERATTACKS. THEY LAUNCHED IT ANYWAY. When OpenAI published the system cards for its new GPT-5.6 family on July 9, the documents contained a detail that deserves more scrutiny than it has gotten. All three models in the family, including Luna, the cheapest and least capable tier, are classified as High risk in both the biological and chemical weapons category and the cybersecurity category. That designation is OpenAI’s own internal assessment. It means the company’s safety team concluded these models are capable of providing meaningful, non-trivial assistance to someone attempting to synthesize a biological agent or execute a serious cyberattack. The only reason OpenAI considers public deployment acceptable is that it has installed safeguards it believes are sufficient to block that assistance in practice. That is a significant thing to bet on. Safeguards get bypassed. Jailbreaks arrive within days of every major model release. The history of every previous GPT launch includes a documented record of safety measures being circumvented by users who put in the effort. OpenAI knows this better than anyone. The models launched anyway. They completed a government security review that the Trump administration declared finished. They are now live for every subscriber on earth. The safeguards are doing the job that the underlying model, by OpenAI’s own accounting, cannot be trusted to do on its own. That should be the headline. It mostly wasn’t. Keywords: GPT-5.6 safety risk, OpenAI bioweapons risk, AI safety card, AI risk assessment, frontier model safety