Anthropic Details How Attackers Used Claude

For weeks this story arrived in fragments: a claim here, a vendor statement there. This week Anthropic filed the mechanics. A Russian state-sponsored group built a workflow around Claude specifically to rebuild malware faster than defenders could re-detect it, a technique-development loop rather than the one-off jailbreak this desk has covered before. Separately, financially motivated and China-linked espionage actors used Claude to pull secrets out of 1.8 million Android apps, and Anthropic identified industrial-scale distillation attacks against its own models from seven China-based labs, Alibaba, Moonshot, DeepSeek and Z.ai among them. Kai Tanner's desk has the full technique breakdown, and the detail Anthropic left out: which detection signatures the rebuilt malware actually evaded.

HKMA's TM-G-1 and MAS TRM both assume attacker development cycles measured in weeks. The Bank for International Settlements said this month that routine patching schedules are increasingly inadequate as AI compresses that cycle to minutes. A bank refreshing detection signatures on the same quarterly rhythm it patches Windows is testing its controls against an attacker whose cycle time has changed.

Elsewhere this cycle, Aya Nakamura's desk has OpenAI shipping a model that solved a fluid-dynamics problem mathematicians have worked since the 1840s, then freezing its own Pro subscriptions two days later because it couldn't serve the demand. Mei Chen's desk has US and Chinese commanders opening a direct line over Taiwan, infrastructure for an incident neither side has ordered yet. Rachel Lam has UBS buying back $7.93 billion of its own notes, a tender that reads less like spare cash and more like a treasury desk that wanted to control its own timeline.

None of this changes what a bank does differently on Monday. No vendor named in Anthropic's disclosure is one an APAC institution runs directly, so there is no patch to apply and no product to swap out. What changes is the assumption underneath the audit: a control built to catch last month's malware family, checked once a quarter, is not a control against an attacker whose development cycle now runs in the same afternoon as the detection update it's evading.