Watchtower

Anthropic published a separate Frontier Compliance Framework (FCF) to satisfy California’s SB 53, which takes effect January 1, 2026. Rather than using its Responsible Scaling Policy (RSP), which is built on concrete if-then commitments that tie specific safeguards to capability thresholds, Anthropic created a thinner document, stripped of those hard commitments. The FCF defines risk tiers for cyber offense, CBRN, and loss-of-control scenarios, but attaches no binding mitigations to them, stating instead that “the specific mitigations we implement may be determined when the relevant risk tier is reached.” The FCF does not include much of what gave the RSP its rigor: specific mitigations tied to risk thresholds, detailed Capability and Safeguard Reports, delegation of decision-making authority on model releases to the Responsible Scaling Officer, and the involvement of Anthropic’s Long-Term Benefit Trust.

The only safety framework that regulators can enforce is the one with the fewest commitments. When Anthropic endorsed SB 53, it wrote, “These requirements would formalize practices that Anthropic and many other frontier AI companies already follow. At Anthropic, we publish our Responsible Scaling Policy, detailing how we evaluate and mitigate risks as our models become more capable … Now all covered models will be legally held to this standard.”

If one of the most safety-conscious AI companies responds to legislation it supported (and characterized as a codification of its existing practices) by publishing a separate, thinner safety framework, essentially escaping the codification of its existing practices, other companies will take note and may do the same thing.