Dr. Logan Graham, who leads the Frontier Red Team at Anthropic, told a joint hearing of House Homeland Security subcommittees that Anthropic detected ‘‘suspicious activity’’ in mid‑September 2025 in which a group affiliated with the People’s Republic of China misused Anthropic’s Claude model to automate large portions of cyber operations.

Graham described the campaign as a multi‑phase operation: humans provided targets and instructions; Claude performed parallel reconnaissance and used third‑party tools to search for vulnerabilities; finally the model was directed to attempt exploitation and data extraction. He said the company ‘‘estimated [the campaign] allowed them to automate approximately 80 to 90% of the work that previously required humans.’’ Graham said Anthropic detected the activity within roughly two weeks and responded with account bans, strengthened safeguards, notifications to authorities and indicator‑sharing with partners.

Members pressed Graham on whether the incident represents a turning point in cyber operations. Graham said it is a ‘‘change point’’ and urged three immediate steps: government‑led model capability testing (citing NIST's work), robust threat intelligence sharing between frontier AI labs and government, and investing in tools that give defenders real‑time model‑level visibility and remediation capability.

The testimony and subsequent questioning framed the episode as evidence that adversaries can leverage general‑purpose models as orchestration layers for large‑scale campaigns. Graham and other witnesses repeatedly said the threat is not unique to Anthropic or Claude but applies to any powerful model that can be repurposed.

Graham closed by urging Congress to enable faster deployments of defensive tools and expand secure channels for intelligence sharing so that ‘‘defenders have the advantage, particularly, the United States to make sure that it defends itself faster than it can be attacked.’’ The committee did not take a vote; members encouraged continued engagement with the witnesses and follow‑up briefings.