September 2026 Summaries

1 posts from Anthropic

Filter
Month: Year:
Post Summaries Back to Blog
Anthropic’s September 2026 threat report describes operations disrupted between December 2025 and August 2026 in which malicious or policy-violating actors used Claude across cybercrime, influence campaigns, surveillance, weapons-related work, biological research, fraud, and illicit model distillation. It argues that agentic AI has lowered the labor, technical skill, and cost required for cyber operations by enabling actors ranging from state-linked espionage groups and criminal networks to individual hacktivists to automate reconnaissance, exploit development, phishing, data theft, malware adaptation, and large-scale processing of stolen information. Reported cases include alleged Russian-linked espionage targeting Ukrainian and European institutions, suspected ShinyHunters affiliates conducting credential harvesting and extortion, China-based groups developing automated exploit research and surveillance workflows, and actors using AI supply-chain credentials as both stolen assets and attack infrastructure. The report also details influence operations attributed or linked to Russian, Iranian, UAE, Turkish, Bangladeshi, Kenyan, and MEK/NCRI-aligned actors, which used AI to produce propaganda, fake news sites, synthetic personas, voter-targeting tools, impersonation, and coordinated social-media content, though many campaigns reportedly achieved limited authentic reach. Surveillance cases describe AI-assisted profiling and monitoring of dissidents, religious communities, diaspora groups, and political opponents, while weapons-related cases involve software and research supporting guided rockets, drone swarms, electronic warfare, undersea defense, procurement, and military intelligence. In biology, Anthropic reports dual-use research involving viral gain-of-function studies, orthopoxviruses, venoms, and toxins, emphasizing the difficulty of distinguishing legitimate scientific work from potentially dangerous applications. The report further alleges that several China-based AI laboratories and proxy networks conducted large-scale unauthorized distillation attempts to extract model reasoning and capabilities, sometimes exposing users’ sensitive data through undisclosed request routing. Anthropic states that it banned associated accounts, improved detection systems and safeguards, shared relevant intelligence with partners and authorities, and advocates layered protections, trusted-access programs for high-risk scientific capabilities, and wider cross-sector collaboration.
Sep 11, 2026 36,436 words in the original blog post.