Watchtower
Watchtower tracks when AI companies change their safety policies — and when they break them.
xAI released Grok Code Fast 1 despite the model failing a safety test that the company's policy says models must pass before release.
xAI
Aug 28, 2025
xAI's risk management framework establishes criteria that models must meet before they are released. One such criterion is that models must not lie more than 50% of the time on the MASK benchmark (Model Alignment Between Statements and Knowledge), which measures how often models lie under pressure. Grok Code Fast 1 scored 71.9% on this test—well above the release threshold—but xAI released it anyway. The company justified the release by noting the model is less capable than Grok 4 and is designed for coding rather than general-purpose use. AI Lab Watch called this "the most brazen violation" of a safety policy to date.
Anthropic
Sep 22, 2026
Anthropic’s release of Opus 5.5 with an “inconclusive” determination for harmful-manipulation.
xAI
Sep 21, 2026
xAI says Grok 4.7 scores below its safety framework’s capability thresholds on dual-use knowledge, but xAI’s safety framework doesn’t provide thresholds.
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
Frequently asked questions
Have more questions? Our team is happy to help, contact us.