xAI quietly dropped two quantitative safety thresholds from its Frontier AI Framework

xAI revised its Frontier AI Framework on June 30, 2026 without announcing it, cutting two numeric risk limits from the December 2025 version — a cap on MASK-benchmark dishonesty and a bar on answering dangerous bio/chem queries on a SecureBio benchmark — along with whistleblower protections and an anonymous reporting channel. Losing the numbers is the substantive shift: xAI's safety claims can no longer be checked against a fixed, measurable line. Zvi Mowshowitz, citing a diff from The Midas Project, flagged the removal as concerning.

Source: thezvi.substack.com ↗