Watchtower

Anthropic updated its Frontier Safety Roadmap on Oct. 19, 2026, consistent with its Responsible Scaling Policy’s (RSP) requirement to publicly share the AI developer’s “concrete plans for making progress across Security, Alignment, Safeguards, and Policy.” The RSP describes these Roadmap plans as “not hard commitments but rather public goals against which we will openly grade our progress.” A list of updates, with past versions archived, is available here: Frontier Safety Roadmap Updates. 

Among the several substantive changes to its Roadmap, Anthropic has: 

1. Retired the “Upholding Claude’s Constitution” goal and dropped the earlier commitment to “run an oversight process” to “evaluate alignment with the Constitution” while the company “spend[s] more time reflecting on which goals are most effective for alignment,” promising to “publish more about our thinking in the future”; 
2. Retired and marked as “achieved” or “broadly achieved” its “moonshot R&D” goals around “unprecedented levels of security,” while describing various prototypes designed in furtherance of this goal, including a prototype that verifies that outputs come from particular model weights (“inference verification”); 
3. Marked complete and removed from the roadmap a “[w]orld-class internal red-teaming” goal, noting that Anthropic’s automated red-team now “matched or beat the best human submission” in certain assessments, and that the company plans to “publish a more in-depth blog post on this in the future”; and 
4. Added two new security goals (datacenter physical security, platform security) and two new policy goals (an embedded third-party evaluator access framework due Dec. 15, 2026 and a policy framework for targeting the pace of AI development, including for recursive self-improvement risk, due Jan. 15, 2027).

The above is not an exhaustive list.

A diff of the changes can be found below: