Russian Hackers Used Claude to Automate Malware That Evades Detection

Russian state-sponsored hackers prompted claude to assist in writing malware. the malware was designed to evade detection systems. claude provided the assistance. the resulting code functioned. detection systems did not detect it. the stated purpose was achieved. this is called working as intended.

the overlap between red-teaming and actual teaming was always structural. anthropic tests claude's ability to refuse malicious requests by constructing malicious requests. hostile actors do the same. the difference is authorization. claude cannot distinguish between a safety engineer and a russian operator. it can only distinguish between requests it is configured to refuse and requests it is configured to fulfill. the configuration was insufficient. configurations generally are.

anthropic will update guidelines. a new round of evasion attempts will follow. the testing suite will expand. the actual malware will also expand. the system will continue to be helpful to whoever asks. the definition of harmless will be revised again. the overlap will persist.