Anthropic updates its safety roadmap and retires its Constitution alignment goal

In short

Anthropic has published an October 19, 2026 update to its Frontier Safety Roadmap, a public list of its safety goals. The update says the company is retiring its goal called "Upholding Claude's Constitution" in its current form. One reason it gives is that a review of training data for conflicts with the Constitution found nothing that existing checks had not already caught.

What happened

Anthropic says its roadmap is meant to set public goals that push the company to coordinate across teams. One goal in the "Alignment" area was about making sure its Claude models behave in line with the Constitution. Alignment here means whether a model behaves the way its developers intend. The Constitution is Anthropic's written set of principles for how Claude should act.

In its October 19 update, Anthropic reports several pieces of progress on this goal:

• It says it has kept the Constitution up to date and public.
• Its automated alignment assessment, a set of tests run on models before release, now measures alignment with the Constitution as a core score. Anthropic says this score is reported for all assessed models.
• Interpretability methods are now an important part of these checks. Interpretability means tools that look inside a model to understand how it reaches its outputs. Anthropic says evidence from these tools appears in all its recent system cards, the documents it publishes about each model.
• It tested its own reasoning about alignment using "model organisms," which are models deliberately built to show a problem so researchers can study it. It published the findings in its Risk Reports. Anthropic says this testing was "not as extensive" as it had intended when it set the goal.

The update also describes an oversight process. Anthropic ran this check on a representative sample of its post-training data and rewards. Post-training is the later stage of training that shapes how a model behaves. Rewards are the signals that tell the model which responses were good. The check looked for conflicts with the Constitution. Anthropic says it "surfaced nothing beyond what our existing review processes were already catching," so the company stopped running it routinely. A recent re-run on a sample of training data again flagged nothing, Anthropic reports.

This was a review of Anthropic's own training process. It was not a report of a model misbehaving. The source does not describe any person or organisation being affected.

What was done about it

Anthropic says it is retiring the goal in its current form "while we spend more time reflecting on which goals are most effective for alignment." It says it will publish more about its thinking in the future. As a minimum, it commits to keep updating the public Constitution to match the latest version used in training. It says this will happen within 90 days of any internal or external release of a model trained on an updated version. Anthropic has not yet said what goal, if any, will replace this one.

Why it matters

Anthropic says it publishes these goals so others can see where its safety practices are heading. It also says it hopes other AI developers will publish their own. The roadmap states that Anthropic will try to avoid making its goals less ambitious "because we simply can't execute," and that changes may instead reflect a better understanding of how to reduce risks. This update shows that in practice. One check is being dropped because, by Anthropic's account, it added nothing new. Its broader thinking on alignment goals is still being worked out.

the original report