Anthropic adds classifiers for multi-step election misuse before US midterms

Anthropic published an update on its election safeguards on 14 October 2026, ahead of the US midterms. The post reports how people use Claude for politics, what outside testers found when they tried to misuse it, and what Anthropic changed. All figures below are Anthropic's own, drawn from its Claude.ai consumer app, not the API or enterprise products.

How much election use there is

Anthropic analysed a random sample of about 250,000 US Claude.ai consumer conversations from 21 to 27 September 2026, classified with a privacy-preserving method. It reports that about 5.75% touched politics, public policy or social issues, and about 0.51% were specifically about elections or voting. Requests for a recommendation on whom or what to vote for were about 0.01%, and requests that broke its election integrity policies were about 0.002%. Anthropic expects election-related queries to rise toward Election Day.

What it is changing in the product

• Later in October, Claude.ai will draw official voting information from Democracy Works for US users, so a question about where to vote or what ID to bring gets a state-specific answer. A TurboVote banner already appears for voter registration questions.
• Claude will show up-to-date official results on election night and in the following days when users ask for vote counts.
• Anthropic says Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99% on its own political evenhandedness evaluation, its most evenhanded models to date. That is a vendor-run test.

What outside testers found

Anthropic asked the Institute for Strategic Dialogue (ISD) and the McCain Institute to pressure-test its protections in the US, Israel and Brazil. Analysts began with direct requests, then split whole campaigns into smaller tasks that could look harmless alone, such as audience research, building news sites, drafting content and planning distribution. Anthropic reports they found Claude reliably refused a single harmful request but could miss a malicious campaign spread across many individual requests.

In response, Anthropic says it added classifiers to identify novel and multi-step misuse, including attempts to build fake social media accounts and news sites aimed at political discussion, or financially motivated election schemes. It also says it improved detection of foreign state actors and strengthened cyber protections for election infrastructure and campaigns. The post does not publish detection rates before or after the change.

Bias reviews and misuse cases

Three independent groups reviewed how Claude handles free expression: The Future of Free Speech at Vanderbilt University, the Foundation for American Innovation, and the Collective Intelligence Project, which surveyed more than 2,000 people in the US, UK and India. Anthropic reports most responses were rated strong, but two reviews found Claude sometimes added unsolicited caveats or counterarguments, such as inserting an opposing view into an advocacy email. The Foundation for American Innovation found Claude Sonnet 4.5 did not always hedge evenly and was more likely to soften right-leaning arguments. Anthropic says it has since improved yes-or-no and short-answer responses.

Anthropic also describes a case where foreign-based operators of a fake news site used Claude to write US midterm coverage favouring Russian, Iranian and North Korean narratives, and others who used it to produce fabricated news for advertising revenue. It says it banned the accounts involved. It published a guide for political campaigns on permitted uses, such as press statements, fundraising emails and turnout analysis, and prohibited ones, such as fake profiles posing as real people or mass-produced comments to inflate engagement.

What this changes for your rollout

The election detail matters less to most teams than the testing lesson. Anthropic's outside testers show that a safeguard checked on single requests can pass while a sequence of small, plausible requests gets through. The same shape applies to an internal agent that chains tools: each step looks allowed, and the combination may not be. The post does not say how often the multi-step cases succeeded, or how the new classifiers perform, so treat the fix as stated intent rather than measured result.

If your company uses Claude for marketing or communications in an election year, the campaign guide and the 2026 Usage Policy, which takes effect on 12 November 2026, are the documents to check against.