The Pulse
Anthropic Retires an Alignment Goal and Adds Two Policy Targets
Anthropic’s October 19 roadmap update reports that an automated red-teaming goal was completed and that a goal concerning Claude’s Constitution was retired in its current form. It also sets deadlines for two new policy goals: one on embedde
AI.info Team · Written with AI · fact-checked ·
Anthropic’s October 19, 2026, update to its Frontier Safety Roadmap reports that the company completed an automated red-teaming goal, retired a goal concerning Claude’s alignment with its Constitution, and added two policy goals. The new goals concern embedded third-party evaluators and the pace of AI development, with deadlines in December 2026 and January 2027.
Automated tests finish ahead of schedule
Anthropic says its automated red-teaming system completed a goal originally due January 1, 2027. The company compared the system with bug-bounty submissions against its biological-weapons classifiers. Using eight bounty questions and three generations of production classifiers, the system matched or beat the best human submission in all 24 question-classifier pairs, winning 23.
The automated tests took 24 to 48 hours, compared with weeks to months for the bounty process, according to Anthropic. The company judged the results more substantial in aggregate than its bounty findings, the standard it set for completing the goal, and said it plans to publish a fuller account later. The reported results cover those specified questions and classifier generations.
Constitution goal ends in its current form
Anthropic also retired its goal to improve oversight of Claude’s alignment with its Constitution. The company says its automated alignment assessments include Constitution alignment as a core metric for all assessed models, and that evidence from interpretability methods appears in recent system cards. It also says it tested its alignment arguments with model organisms and published findings in Risk Reports, but did so less extensively than the original goal envisioned.
The company initially reviewed a representative sample of post-training data and rewards for alignment with the Constitution. Anthropic says that review found nothing beyond what its existing processes caught, so it stopped running the process routinely; a later sample review also flagged nothing. The retirement applies to the goal in its current form. Anthropic says it plans to continue updating the public Constitution to match the version used to train models within 90 days of relevant deployments.
Two policy goals receive deadlines
The first new policy goal is a proposal on embedded third-party evaluators, due December 15, 2026. Anthropic says the proposal will address how governments can support deeper access for independent evaluators, including funding, the access they need, and protections for their independence. The company says it has committed to giving evaluators deeper access to models and internal information while preserving their independence, and that it will draw on its experience working with evaluators.
The second goal, due January 15, 2027, is a policy framework for targeting the pace of AI development. Anthropic says its current policy framework focuses on blocking or deterring the deployment of dangerous AI systems. Its new policy work will consider options for addressing risks from the speed of development, particularly if automation of AI research and development accelerates progress faster than governance or safety measures can keep up. The company says it will publish a proposal or explain why it decided not to.
A broader policy revision
The roadmap changes sit within Anthropic’s broader Responsible Scaling Policy revision, Version 3.0, published February 24, 2026. In its announcement, Anthropic said the revised policy separates measures the company plans to pursue regardless of what other organizations do from a broader map of mitigations it believes the AI industry should adopt. The announcement also described public roadmaps and recurring Risk Reports as part of the policy’s transparency and accountability measures.
The October update records several different decisions: Anthropic reports completing the red-teaming goal, retiring the Constitution goal in its current form, and setting deadlines for two new policy goals. The roadmap describes its goals as public priorities and says it will provide updates on whether they are achieved, with new goals set when that happens.