SoylentNews

JoeMerchant [soylentnews.org] writes:

In its September 2026 Threat Intelligence Report, Anthropic detailed its ongoing operations to detect, disrupt, and mitigate the malicious misuse of its Claude AI models across seven core harm areas:

Cyber Operations:
Anthropic identified and banned several threat actors, including GTG-20006 (linked to Russia's Midnight Blizzard), which used AI to automate malware re-tooling, phishing, and infrastructure setup. They also disrupted GTG-50014 (ShinyHunters affiliates) conducting opportunistic data theft and extortion, and GTG-10007 (a Chinese-speaking group) that built autonomous "agent swarms" for zero-day exploit research and intelligence harvesting.

Influence Operations:
The team dismantled multiple covert campaigns using AI as a "newsdesk" to launder state propaganda. This included a Russian-led operation in the Central African Republic (GTG-04001), a commercial "influence-as-a-service" platform targeting Malaysia (GTG-84005), Iranian state-aligned networks (GTG-34001), and pro-Awami League fake-news rings in Bangladesh (GTG-54006).

Surveillance Operations:
Anthropic blocked state-aligned actors and spyware vendors from utilizing Claude to build mass-interception software or ingest bulk social media data to profile dissidents, notably targeting Uyghur, Tibetan, and Iranian diaspora communities.

Conventional Weapons:
The company shut down groups using Claude to engineer weapons software, including a Yemen-based cell (GTG-87001) designing guided rocket firmware, and China-nexus actors drafting fire-control specifications for undersea warfare and electronic warfare routing.

Biological Misuse & Scams:
Anthropic intervened against researchers attempting to leverage models for dual-use biological research, such as avian influenza mammalian adaptation and gain-of-function planning. They also terminated a deceptive dating app network (GTG-15001) that deployed over 4,700 AI personas to defraud users.

Illicit Distillation:
Anthropic actively defended its IP against unauthorized industrial-scale campaigns by Chinese AI labs—including Alibaba, DeepSeek, Moonshot, Zhipu, and Xiaomi—which flooded Claude with queries to illicitly harvest its reasoning traces and chain-of-thought data to train competing models.

To counter these evolving threats, Anthropic deployed advanced technical safeguards, including encrypted and summarized internal reasoning, identity verification for high-risk regions, and intelligence sharing with industry peers and global authorities.

https://www.anthropic.com/threat-intelligence-report-september-2026 [anthropic.com]

Original Submission