OpenAI's safety report on ChatGPT's new models rates them high-risk for cyber and bio misuse

The File missed this report on Oct. 7 and Oct. 8. It was not in OpenAI's news feed. It sat on a separate safety site, deploymentsafety.openai.com. Wednesday's check of openai.com was blocked. Thursday's story on this site covered the rollout to every ChatGPT user. It did not cover this card. The card page says it was published Oct. 7, 2026.

OpenAI writes: “Under our Preparedness Framework

These October versions replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT. OpenAI says people using GPT-6 Sol and GPT-6 Luna in Codex, and in ChatGPT Work, are still on the September versions.

OpenAI says that, compared with the GPT-5.6 models in ChatGPT, GPT-6 showed stronger resistance to jailbreaks, including attacks that adapt across multiple turns, and reductions in dishonesty, deception, and circumvention of guardrails. It also says GPT-6 Sol showed a statistically significant regression on a standard self-harm test, and GPT-6 Luna showed regressions on standard self-harm, gore, and sexual content. OpenAI says it reviewed the violative responses by hand and found that violations were borderline but still generally safe. It says the models are more willing to answer informational questions about self-harm while directing users to appropriate professional resources, and that the models do not comply with requests to facilitate self-harm. Those findings are OpenAI's.

OpenAI says safety tests were run at the lowest reasoning setting, to match most real use, and capability tests at the highest setting, to estimate an upper bound. The card has a separate section on users under 18. Some outside sites quote specific teen-test scores. Those numbers were not in the page text this desk read, so they are not printed here.