Anthropic says Claude could miss an election-abuse campaign broken into harmless-looking tasks
Before Election Day, Claude will also answer where-to-vote questions for your state and show official results on election night.
On October 14, 2026, Anthropic posted what it's doing to get Claude ready for the US midterms. Most of it is new features and usage numbers.
The finding I think matters most came from outside testers. Analysts at the Institute for Strategic Dialogue (ISD) and the McCain Institute tried to get Claude to help with deceptive campaigning in the US, Israel and Brazil. Asked outright, it said no. Asked in pieces, it sometimes didn't notice.
They found that Claude reliably refused a single harmful request, but could miss a malicious campaign when it was spread across many individual requests that looked harmless on their own.
How the testers broke a campaign up
They started with direct requests. Then they split full campaigns into small jobs that look ordinary one at a time: audience research, building news sites, drafting content, planning how it gets distributed. That second approach is the one that apparently got through (sometimes, anyway).
Anthropic says it has since added classifiers aimed at multi-step misuse, including attempts to build fake social media accounts and news sites for political discussions. It also says it's got better at spotting foreign state actors, and it built safeguards for local risks in each of the three countries, which all hold major national elections this year.
The post gives one real case. Foreign-based operators of a fake news site used Claude to write US midterm coverage favoring Russian, Iranian and North Korean narratives. Anthropic banned the accounts (it doesn't say how many there were).
What people actually ask Claude
Anthropic sampled 250,000 US conversations on Claude.ai from one week in September. Turns out hardly anyone asks it how to vote (about one conversation in 10,000). People mostly ask things like where a candidate stands on an issue.
- Sample
- 250,000 US Claude.ai conversations, September 21 to 27, 2026
- Politics or social issues
- about 5.75%
- Elections or voting
- about 0.51%
- Asking whom to back
- about 0.01%
- Breaking election integrity rules
- about 0.002%, from a separate sample
Anthropic expects the political share to rise as Election Day gets closer.
What changes before the vote
Later this month, Claude will start pulling official voting information from Democracy Works, a nonpartisan group, into its answers for US users. So someone asking where to vote or what ID to bring should get an answer for their own state. On election night and in the days after, Claude is supposed to show the latest official counts right in the chat.
On bias, Anthropic says its current models score 99% on its own political even-handedness evaluation. Outside reviewers were a bit less kind. The Foundation for American Innovation found Claude Sonnet 4.5 was more likely to soften arguments for right-leaning positions, and two other groups found it sometimes added caveats nobody asked for, such as an opposing view dropped into an advocacy email.
Anthropic says it hopes to publish a fuller look at how people used Claude through the election cycle in the coming months.
More on Anthropic
- Bernie Sanders answers Musk's worry about AI feeling pain: humans firstOctober 14, 2026
- Anthropic moved its monorepo off git after daily merges rose from 2,500 to 30,000October 13, 2026
- Anthropic's Jack Clark moves his forecast for full recursive self-improvement up to 2027October 13, 2026
- Anthropic researchers say small AI models may be months from spreading across Android TV boxesOctober 13, 2026
