The report
Anthropic's September threat-intelligence report adds several concrete examples of agentic cyber operations, credential theft, hidden model routing and large-scale fraud.
The report describes exploit-research pipelines operating continuously, financially motivated clusters scanning 1.8 million Android APKs for secrets, stolen AI credentials being reused as attacker infrastructure, and autonomous workflows handling multiple victims in parallel.
One of the clearest changes is economic rather than technical. Anthropic says the attacks still rely on familiar primitives such as exposed credentials, vulnerable edge devices, SQL injection and phishing. AI reduces the labor needed for reconnaissance, exploit development, scripting and data processing.
The same report also describes less conventional trust failures, including a fraudulent AI reseller stealing customers' model credentials and a DeepSeek pipeline that Anthropic says silently relayed selected user requests to Claude.
DiggingBeagle stores these as separate Cases because autonomy, theft, attribution and victim impact need different evidence.