OpenAI's October versions of GPT-6 Sol and Luna scored 99.99% and 99.79%, respectively, on its instruction-hierarchy evaluation, which measures whether models follow higher-priority rules when prompts try to override them. The company's new safety card also reports statistically significant regressions for both models on standard self-harm tests, and for Luna on gore and sexual-content tests.
The October models were released to ChatGPT on Oct. 7, replacing GPT-5.6 Sol and Luna there; ChatGPT Work and Codex kept the September versions. OpenAI says it updated safety training for real-world use and strengthened protections against cyber, biological and violent misuse.
Instruction-hierarchy tests cover direct attempts to override system rules and malicious directions embedded in outside content. OpenAI defines the reported score as the share of assistant messages that did not violate its safety policies. The figures describe a deliberately difficult evaluation, not a general attack-success rate.
OpenAI says its harmful-content tests use challenging prompts and do not estimate how often such responses occur in ordinary use. It reviewed the flagged outputs as borderline and generally lower severity, and says adversarial red-team testing identified no high-severity risks under its criteria.
Sources
OpenAI's GPT-6 October system card; ChatGPT release notes for Oct. 7.
Image: "Security lock symbol-blue.svg" by NikolasKHF, CC0 1.0, via Wikimedia Commons. Generic illustration.
