Watchtower
Watchtower tracks when AI companies change their safety policies — and when they break them.
OpenAI updated its system card for two of its models: GPT 5.6 and GPT Live.
OpenAI
Aug 3, 2026
Between August 3 and August 4, 2026, OpenAI updated its system card for two of its models: GPT 5.6 and GPT Live.
On August 3, they changed the GPT-5.6 model card to add performance on a new red-teaming approach called GPT-Red. The updated model card has been timestamped and the newly-added content is reflected as new; you can read the full model card here.
On August 4, 2026. OpenAI added the following disclaimer to their model card for GPT Live:
“We identified a configuration mismatch in the safety evaluation setup: the results had been generated using a backend configuration that did not match the final model release. We re-ran the full evaluation using the correct configuration and updated the reported results. Some values remained unchanged after rerunning, while others were corrected. We’ve added an appendix to show the original and corrected values side by side. The overall safety conclusions are unchanged; none of the regressions led to numbers below our safety launch standards.”
The current GPT Live model card, with the added comparison appendix, can be found here.
OpenAI
Sep 9, 2026
Updated the GPT-6 Astra system card, specifically to hedge claims about misalignment.
OpenAI
Aug 18, 2026
OpenAI updated its Model Spec, the document outlining intended model behavior
xAI
Aug 17, 2026
xAI updated Grok 4.6's model card after release. A changelog was included.
Aug 14, 2026
Google updated its Gemini 3.7 Flash model card after publication, making changes to language in the “Key Results for Gemini 3.7 Flash” column in the Frontier Safety Assessment section
xAI
Jul 20, 2026
xAI made changes throughout the model card for Grok 4.5, including some that appear to be persistent errors
Frequently asked questions
Have more questions? Our team is happy to help, contact us.