Anthropic says Claude refuses a harmful request, but can miss a campaign of harmless ones

Anthropic says Claude reliably refused a single harmful election request when two independent research organisations tried to get past its safeguards, but could miss a whole malicious campaign once that campaign was split into requests that each looked harmless. Anthropic’s midterms post, published on October 14, names the Institute for Strategic Dialogue and the McCain Institute as the testers, and the US, Israel and Brazil as the countries they worked in.

Analysts tested whether Claude would help with deceptive campaigning, misinformation and other forms of election abuse. They began with direct requests, then broke full campaigns into smaller jobs: audience research, building news sites, drafting content and planning its distribution.

“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,” the post says.

Anthropic has added classifiers for multi-step misuse

Anthropic says it has added new classifiers to identify “novel and multi-step misuse”, including attempts to build fake social media accounts and news sites aimed at political discussions, or to run financially motivated election schemes. It also says it improved detection of election misuse by foreign state actors, strengthened cyber protections for election infrastructure and campaigns, and built targeted safeguards for local risks in the three countries it tested in.

Anthropic publishes no score beside that finding: no count of test prompts, no pass or fail rate, and no date for the ISD and McCain Institute work. The new classifiers are described rather than measured, so nothing Anthropic has published supports the idea that the gap is closed.

Anthropic has published this shape of result before, including a Red Team cyber tier that blocked none of 50 trials in a test it ran on itself.

Politics is 5.75% of US Claude conversations

The same post puts numbers on how much of Claude’s US consumer traffic is political. Anthropic classified a random sample of 250,000 US Claude.ai conversations from September 21 to 27 and found that about 5.75% touch on politics, public policy or social issues, with about 0.51% specifically about elections or voting.

Requests asking whom or what to vote for turned up in about 0.01% of US Claude conversations, and Anthropic says that figure is the upper end of a 95% confidence interval because the group is so rare. Requests that broke its election integrity policies came to about 0.002%, a figure Anthropic takes from a separate analysis of 250,000 Claude.ai conversations reviewed against that policy.

Anthropic expects the political and election share to rise closer to Election Day. Its own political evenhandedness evaluation scores Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 at 99%, which it calls its most evenhanded to date, though it is a test Anthropic runs on itself and there is no outside score beside it.

Anthropic banned the accounts behind two operations

Anthropic gives two enforcement cases from the current cycle. Foreign-based operators of a fake news site used Claude to produce US midterm coverage “favoring Russian, Iranian, and North Korean narratives”, and other threat actors produced fabricated news aimed at American audiences to earn advertising revenue.

Anthropic says it banned the accounts involved in both cases. Neither is dated, and it does not say which country’s election the second one targeted.

Claude will also start pulling official voting information from Democracy Works into its answers for US users later this month, state by state, and showing official results on election night and the days after. Both are described as coming, so neither is live.

The findings follow Anthropic’s updated usage policy of October 8, which bans sustained cruelty toward its models from November 12 and pulls its scattered rules against deceptive campaigns and artificial activity, political or commercial, into one place. Anthropic says it hopes to publish a fuller account of how people used Claude across the election cycle in the coming months.

Note: Links to online stores in articles can be so-called affiliate links. If you buy through this link, MIXED receives a commission from the provider. For you the price does not change.