Anthropic's New Rules for Claude, in Plain English

Anthropic updated its usage policy on Oct. 8, 2026, for the first time in over a year. The changes cover election interference, weapons development, surveillance, and health and financial uses. The Verge

The company describes the Usage Policy as a rulebook for how Claude should and should not be used. Anthropic Its newsroom lists a '2026 Usage Policy update' announcement dated Oct 8, 2026 under Announcements.

The revision pulls several separate risk areas together into clear bans. It also adds a narrow rule about how users treat the model itself. For that rule, the model itself handles enforcement.

Elections and deceptive campaigns

Anthropic added a ban on dishonest commercial or political campaigns. This covers hiding who is behind a message or boosting content with fake accounts or posts.

Separate election rules ban tricking voters and disrupting elections. That includes false claims about candidates or how to vote, pretending to be candidates or election officials, or stopping people from voting.

The rules focus on who sent a message and how far it spreads. They do not ban political persuasion itself. They ban hidden sponsors and fake amplification, and false claims about voting rules that keep people from voting.

Weapons and physical autonomy

Anthropic widened its weapons ban to include software and parts that make weapons work, including putting weapons on drones and other self-moving vehicles.

The detailed policy bans building or running software or parts for testing or using weapons, including aiming, fire control, or attack. It bans putting weapons on drones, vehicles, or any unmanned or self-moving platforms, or building weapons delivery systems.

The wider ban covers developing, making, changing, designing, or testing weapons, explosives, dangerous materials, or weapons functions, including altered hardware built to kill or destroy. It separately bans work on biological, chemical, radiological, or nuclear weapons, explosive or homemade bombs, or starting materials or delivery methods, including key weapons technology such as making fissile material, enrichment, or bomb design. It also bans changing or testing biological or chemical agents to make them more lethal, more contagious, more harmful, longer-lasting, or harder to detect or treat. Illegally getting, selling, or moving weapons or getting around licenses or export controls is also banned.

A related safety rule covers machines that can act in the physical world. When a model is connected to equipment that can move on its own and cause injury, a trained operator must be able to watch the equipment and stop it if needed. This applies no matter the intent. A person must stay in charge, like a supervisor with an emergency stop button.

Surveillance, tracking and criminal justice

Anthropic bans tracking people without their consent, whether live or by studying data collected earlier. It bans using Claude to build or improve tools made for surveillance.

For police and courts, the line is at targeting individuals. Anthropic bans using Claude to decide or suggest who to investigate, arrest, or charge.

The policy says its rules can be changed for contracts with certain government customers if the contract has strong limits and safety steps to reduce harm. That exception is conditional. The contract must include protections.

Related hacking rules ban breaking in and building break-in tools. They include finding or using security holes without the owner's permission, getting in without permission or gaining higher control through technical attacks or tricking people, building tools to intercept messages or watch devices without permission, building lasting secret remote-access tools including changes to low-level hardware software or hardware implants, building automatic tools to break into many systems at once without permission, getting around logins or security software without permission, making malware or ransomware, and running denial-of-service attacks that shut systems down by flooding them or running botnets, which are networks of hacked computers. The same section bans helping destroy critical systems such as power grids, water treatment plants, medical devices, phone and internet networks, emergency services, or transport systems, breaking into election, health care, or financial systems without permission, and interfering with military bases and related sites.

There are allowances for defense. The ban does not stop security research, testing, or tool building on systems the user owns or runs, with the owner's permission, or in an approved bug bounty or reporting program, if it follows the law. Users covered by real-time cyber safeguards can ask for adjusted access through its Cyber Verification Program.

Other standing bans remain, including illegally making, getting, selling, or sharing controlled drugs, taking part in or helping human trafficking or prostitution, and violating other people's intellectual property rights.

Behavior toward the model

The updated policy bans "sustained and needless abusive or cruel behavior" toward Claude.

This follows Anthropic's August research note that Claude Opus 4 and 4.1 can end a rare set of conversations in rare, extreme cases of ongoing harmful or abusive user behavior. Anthropic The company described that ability as part of research into "model welfare."

Ending the conversation remains the "primary enforcement mechanism" for abuse toward Claude. Anthropic said the rule applies only in extreme cases of repeated cruelty with no clear purpose. It does not apply to user frustration, pushback, dark creative themes, or model testing and research.

The limits are specific. Venting after a bad output is allowed. Red-teaming, which means testing to find safety flaws, is allowed. Fiction with violent or disturbing material is allowed. What is banned is long, cruel behavior for its own sake.

The broader context here is that usage policies are now instruction manuals, not only legal papers. They set when the model should refuse, what tools it can use, and when a person must stay in control.

In my view, the biggest change is the move from single answers to wider effects. The concern is less about one answer and more about what that answer allows when linked to drones, tracking systems, or influence campaigns. That matches how Claude is actually used, as a part inside bigger systems rather than as a chatbot answering single questions. The exception for government contracts will get close attention, because uses blocked for most users can be allowed again under contract protections. The test will be how Anthropic defines good enough protection in practice. If the policy holds, developers building on Claude should get clearer refusal rules, and stronger human control where models touch the physical world.