Watchtower

Anthropic updated their Frontier Compliance Framework (FCF), the compliance-facing document that serves as Anthropic's framework under California's TFAIA and the EU General-Purpose AI Code of Practice (EU CoP). This is the second revision since the FCF's December 2025 release, following the March 2026 update.

One substantive change reshapes the CBRN Tier 2 threshold for “Novel chemical/biological weapons production” to match recent updates in Anthropic’s Responsible Scaling Policy (RSP), which we covered previously. In short, the capability threshold was narrowed while the harm it covers was broadened, and Anthropic asserts that the original and v3.3 wordings reach the same conclusion.

The FCF also expanded the “Sabotage and loss of control” Tier 2 threshold for automated R&D. This update adds an explicit definition of when the threshold is met: either (1) a model that can fully substitute for Anthropic's entire set of Research Scientists and Research Engineers at competitive cost (within a factor of 5), or (2) "dramatic acceleration" of AI progress, defined as a 2x rate increase plausibly attributable to R&D automation rather than headcount, compute, or general productivity. The earlier "effective compute scaling" reference was also dropped.

Other changes were minor and included terminology changes to match the RSP and an update to the Table of Contents to include section 2.6 that the March version omitted. Cyber Offense, CBRN Tier 1, and both Harmful Manipulation tiers are unchanged.

Anthropic does not provide a browsable public archive of past FCF versions the way it does for its Responsible Scaling Policy; superseded versions are not surfaced in its trust center.

A diff of the changes can be found below:

‍