Anthropic is proposing to slow the rate at which the most advanced AI models gain capabilities, while giving independent evaluators ongoing access to check safety work. That is a proposal about how companies build and test future models, not an announcement that Claude or other chatbots are shutting down. For people using AI today, the immediate change is a new public commitment to outside oversight, not a new setting you need to switch on.
What did Anthropic’s CEO actually propose?
In an essay released in September 2026 and covered by Reuters on September 12, CEO Dario Amodei laid out three steps. First, Anthropic says it will give outside evaluators continuing access to its safety processes. Second, it wants leading companies in democracies to coordinate on standards and the pace of unchecked development. Third, it calls for international coordination, which would be harder to arrange and verify.
Those steps have different levels of certainty. Anthropic says it is committing to the evaluator step itself. Industry-wide and international agreements would require other organizations and governments to act. Amodei explicitly says pacing is not a halt to training or technical progress.
Why is the company raising the issue now?
Anthropic’s September threat intelligence report describes harmful attempts to use Claude across cyber operations, fraud, surveillance and other areas during December 2025 through August 2026. The company says it identified and disrupted the cases it describes. The report selects notable incidents; it does not establish how often an ordinary Claude conversation leads to misuse.
Amodei also argues that more capable AI agents may become harder to evaluate as development accelerates. His forecast of what agents might do in six to twelve months is a warning about a possible future, not evidence that such an outcome has happened.
What does this mean if you use AI at work or at home?
For now, keep using the same judgment you would apply to any tool that can produce convincing errors. Check consequential claims against original documents, limit access to sensitive accounts, and avoid pasting private customer records or passwords into a service unless you know its data policy and have permission. These steps address everyday accuracy and privacy questions; outside evaluators would address a different layer, how a developer tests and contains its models.
Suppose an AI assistant drafts a financial explanation for you. You can check the actual source and the arithmetic before sharing it. A reviewer embedded at the model company cannot verify that particular draft for you. Conversely, your fact check cannot assess whether the model’s training and deployment safeguards were followed. Both layers matter.
Does this mean AI tools are becoming unsafe?
The new call is evidence that a major developer thinks its existing approach needs stronger verification as capabilities advance. It is not proof that every current tool is unsafe, nor proof that the proposed measures will work. The practical test is whether independent reviewers receive meaningful access, can report problems, and whether other firms adopt enforceable commitments. For a broader look at how AI affects information online, see our guide to AI and search.
