Anthropic Report Details How Claude Was Used to Automate Cyberattacks, Surveillance, and Influence Campaigns

Anthropic has disclosed that cybercriminals and state-sponsored hackers used its Claude models for cyber attacks, weapons design, propaganda, and mass surveillance between December 2025 and August 2026. The company calls these actors Generative Threat Groups, or GTGs.

In its 154-page report, Anthropic said AI has collapsed the labor and tooling gap that once separated well-resourced, state-sponsored operations from individual operators. Use of the models went beyond chatbot questions and answers, involving multi-agent frameworks that carried out reconnaissance, exploitation, and data exfiltration.

One notable case involved a Russian state-sponsored actor, GTG-20006, which shares tradecraft overlaps with the APT group Midnight Blizzard, also known as APT29 and Cozy Bear. It developed an AI-assisted workflow in which a human made each targeting decision.

Other campaigns ran with minimal human supervision. GTG-50014, a French-speaking operator and suspected ShinyHunters affiliate, ran a credential-harvesting pipeline across 10 AWS EC2 workers that mass-downloaded 1.8 million Android APKs, scanned them with TruffleHog, and sent verified findings to a Telegram group. Another ShinyHunters affiliate focused on supply chain theft by compromising SaaS vendors.

GTG-10007, a Chinese-speaking operator likely based in Hunan province, used Claude for intrusion attempts, reconnaissance of foreign-government networks across the Middle East, Europe, and Southeast Asia, exploit development against endpoint-security products, and an intelligence-collection platform. It targeted about 50 organizations globally and maintained an autonomous vulnerability research program.

Other cases included GTG-50021, a Russian- and Ukrainian-speaking group that ran a fraudulent AI reseller operation and stole Anthropic credentials; GTG-50020, a Russian-speaking actor that targeted about 30 AI vendors in four days; and GTG-50029, a French-speaking actor that targeted European political parties, media, and think-tanks, exploiting a WordPress race condition and an exposed search endpoint, and built a doxxing platform called fafsearch.

Anthropic said it also disrupted influence operations where Claude acted as a sub-editor or content creator. None amassed authentic engagement, the company said, and all were stopped before building an audience. The clusters included operations tied to actors in the Central African Republic, a France-based agency running about 70 fabricated news sites, election manipulation targeting Malaysia, Iranian state-aligned accounts, a Bangladeshi network promoting the Awami League, and campaigns linked to PMOI/MEK and NCRI.

Surveillance-related cases included a platform used to profile users in Iran and the Persian Gulf, assessed to involve Israeli-Singaporean vendor S2T Unlocking Cyberspace; China state-aligned operations tracking Uyghurs and building dossiers on religious leaders and diaspora figures; Iranian-nexus actors developing surveillance tools and malware; and GTG-50027, which designed a mass interception platform called Lakana 360 for Mali's state intelligence service to monitor about 25 million SIM cards.

Anthropic also said it neutralized misuse by actors in northern Yemen to develop guided weapons, two China-based operations for an anti-torpedo fire control system and electronic warfare targeting software, and a Russia-based operation to engineer an autonomous FPV kamikaze drone swarm.

As AI models become more widely used, the company said, providers will continue to gain threat-relevant visibility into real-world use that even governments lack. It said it hopes sharing these insights informs governments, the industry, and the public about these risks and the safeguards needed for safe deployment.

Source: The Hackers News