Nov 12: cruelty to Claude could get you banned. The real shift is bigger

Starting November 12, Anthropic will enforce a new usage-policy rule that prohibits "sustained and needless abusive or cruel behavior" toward Claude. The company says the ban applies only in extreme cases, where users repeatedly act cruelly toward its models with no discernible purpose, and it does not cover common versions of user frustration, pushback, dark creative themes, or model testing and research. In other words, you can still tell Claude it is wrong, but repeatedly berating it for the sheer pleasure of doing so could put your account at risk.

The cruelty clause is the attention-grabber, but the same policy overhaul quietly rewrites the rules on far more consequential matters: election interference, weapons development, and surveillance. Anthropic says it tightened restrictions on deceptive influence campaigns after observing state media outlets, government propaganda offices, and commercial organizations using Claude to operate fake accounts and fabricated news websites. At the same time, it dropped its blanket ban on personalized political targeting, arguing that the restriction also caught legitimate activities, such as nonprofits translating voter information. Deceptive targeting and misuse of personal information remain prohibited.

On weapons, Anthropic clarified that its ban extends to software and components after observing attempts to use Claude to develop guidance and control systems for weapons, including armed drones and autonomous vehicles. The company also made clear that Claude cannot be used to recommend which people police should investigate, arrest, or charge, nor to develop tools for surveillance. Tracking people without their consent is prohibited, whether in real time or retrospectively. Anthropic frames these changes as clarifications of existing restrictions rather than new ones, but the specificity suggests real-world misuse cases are already surfacing.

The policy also introduces new requirements for customers connecting Claude to physical equipment capable of causing injury. A qualified human operator must be able to monitor and stop the hardware, and the equipment must remain in a safe state if the AI connection drops. That is a meaningful shift for enterprise deployments in robotics, manufacturing, or any setting where AI output can trigger physical action. It effectively puts a human-in-the-loop obligation on the customer, not just on the model.

On the etiquette front, Anthropic is not claiming Claude experiences distress. The company acknowledges there is no established evidence that the software suffers, yet it is still telling paying customers to mind their manners. The primary enforcement tool will be Claude's existing ability to end abusive conversations, a feature introduced in August 2025 as part of Anthropic's research into what it calls "model welfare." Some versions of Claude can already cut off users who persistently subject the models to abuse. What remains unclear is how Anthropic will distinguish legitimate frustration from gratuitous cruelty, since both can look similar in a chat log.

The policy overhaul comes as Anthropic tries to address the security risks posed by increasingly capable AI systems. The company has launched a new cybersecurity initiative aimed at helping critical infrastructure operators and open source projects find and fix vulnerabilities using its AI models. The effort includes free security scanning for eligible open source projects and partnerships with security firms to help protect critical infrastructure. Anthropic reckons attackers will have the advantage for the next two years before AI-powered defenses begin to catch up.

For now, though, the company has another threat to contend with: people saying nasty things to its chatbot. The November 12 effective date gives users just over a month to get any lingering insults out of their systems. For enterprise customers, the broader lesson is that AI usage policies are becoming more explicit about conduct, oversight, and physical safety. Boards and operators should treat this as a signal that vendors will increasingly hold customers accountable for how AI is deployed, not just what the model outputs.

The same policy that polices chatbot manners also draws harder lines around weapons, surveillance, and physical AI systems. For any board with AI in its operations, that combination is worth watching closely.