Watchtower

OpenAI updated its GPT-6 Astra System Card on Sept. 9, 2026—originally published just six days before. The update includes a two paragraph-long change log identifying changes to the Alignment section (Section 8) and changes to the “naming and substance” of the portion dealing with “Verbalized Metagaming and Oversight Gaming” (Section 8.7).

While OpenAI’s Astra release highlighted that the new model represented a step forward in alignment, the revised system card includes new caveats and clarifications suggesting evidence of Astra’s overall alignment should be read in context of the overall system card, including Astra’s higher evaluation awareness and limits of monitoring.

Additionally, the “naming and substance” revisions impact OpenAI’s analysis of its new “honeypot” evaluation—which OpenAI specially crafted following the Hugging Face incident. OpenAI’s system card changes what it previously referred to as “alignment faking” to “oversight gaming” and casts doubt on whether Astra recognized the “honeypot” and thus avoided it, stating that Astra’s metagaming “seems to be largely about whether the task is intentionally impossible.” OpenAI also clarifies that it “did not construct honeypot training environments specifically to target this evaluation or reproduce the incident.”