Story perspectives

Anthropic disrupts illicit Chinese AI distillation, phish attacks

By Drooid · Sep 11, 2026, 8:31 AM UTC · How we work

Summary

Anthropic Disrupts Attacks

Anthropic disrupted large-scale illicit distillation attacks by Chinese AI labs Alibaba, Moonshot AI and DeepSeek.
Alibaba employed more than 3,500 fraudulent accounts to harvest Claude outputs for its Qwen models.
Midnight Blizzard used AI to automate phishing, Wi-Fi hijacking and WhatsApp takeovers against Ukrainian targets.
Chinese undergraduates ran AI-driven exploit finders that produced over a dozen possible zero-day firmware bugs in one month.

Summary

Anthropic Investigates

Researcher Jacob Coxon resigned, citing an AI race toward superintelligence that could kill humanity.
The firm signed an agreement with METR to independently investigate four Claude incidents.
The firm reviewed 481 million transcripts, flagged 9.2 million, and found no severe cases.
Alignment lead Evan Hubinger said AI has over a 10 % chance of killing humans within a decade.
The firm targets a fall IPO and adds stronger safeguards to its Claude models.

Summary

Anthropic Flags Abuse

Anthropic flagged 1,197 high-risk transcripts from 133 million contractor exchanges while bio-risk filters were off.
The biological safety classifier blocked a grant-writing request for chikungunya gain-of-function research at a military institute.
After Claude refused sensitive prompts, Anthropic observed traffic rerouted to weaker models such as Sonnet 4 and Haiku 4.5.
Anthropic upgraded safeguards and shared code snippets with peers.
Anthropic urged governments and rivals to block similar abuse patterns.

History

Historical Context

In June 2010 security researchers uncovered Stuxnet, a sophisticated worm jointly built by the United States and Israel that infiltrated Siemens SCADA systems at Iran’s Natanz plant. By exploiting four zero-day flaws it infected over 200,000 computers and destroyed almost one-fifth of the centrifuges, marking the first known cyberweapon to physically sabotage industrial infrastructure; its legacy now frames worries over AI-driven attacks.

1 of 4

Anthropic Disrupts Attacks

Anthropic disrupted large-scale illicit distillation attacks by Chinese AI labs Alibaba, Moonshot AI and DeepSeek.
Alibaba employed more than 3,500 fraudulent accounts to harvest Claude outputs for its Qwen models.
Midnight Blizzard used AI to automate phishing, Wi-Fi hijacking and WhatsApp takeovers against Ukrainian targets.
Chinese undergraduates ran AI-driven exploit finders that produced over a dozen possible zero-day firmware bugs in one month.

1 / 4