Anthropic sorts eight months of Claude misuse into seven named harm areas

At least 16 accounts tied to Iranian paramilitary and domestic security agencies used malicious browser extensions built by Claude to harvest data from 6,388 Iranians, according to Anthropic ^[15]. Divide the second number by the first and you get about 399 people per account, and because 16 is a floor, the real average is lower ^[17]. In the same section the company says an Iran-nexus threat actor used Claude to pinpoint U.S. naval targets ^[16].

In the biological cases the failure has two stages. Anthropic said the users "circumvented controls" designed to prevent individuals in countries without access to Claude from using the product, and also attempted to hide the purpose of their research to evade safeguards ^[6]. The company described the individuals in each case study as working scientists and did not name them ^[7]. "We do not assert that they intended harm, and identifying them or their labs could expose them to harm," the company wrote in its report ^[8]. It banned the accounts upon discovering the activity ^[9].

Mashable's account of the report counts five biological case studies ^[4] and nine surveillance instances spanning China, Iran, west African countries and for-hire networks ^[13], which is 14 enumerated cases in two of the categories ^[18]. Anthropic did not include details on how long the activity had been occurring in most cases ^[3]. Those are numerators with no denominator: no count of accounts reviewed, no time between first prompt and ban.

Anthropic argues the record itself is the new thing. "Historically, this kind of work has been uncovered by governments, United Nations panels, and outside investigators, who piece it together from recovered hardware and public sources," the company explained ^[10]. The China-based account in the weapons section fits that description of a gap: Anthropic says the user, possibly a military-industrial researcher, built electronic warfare modules with Claude work tools to defeat enemy radar and communications and suppress air defenses, in a simulation that included 12 targets in Taiwan ^[12]. That came out of account activity on a vendor's servers, and the vendor chose how much of it to print.

The next document of this kind, from any vendor, can be sorted on two axes. Does the item carry a denominator, and was the misuse caught during the session or after the fact. Items with both are a control test. Items with neither are a case list, which is good material for drafting policy and no use for measuring whether the policy works. Almost everything here sits in the second group, which is still more than a vendor safety page gives a buyer. The two figures that would move items into the first group are accounts reviewed and time from first prompt to ban.