Anthropic releases report on AI misuse by criminals and state actors

Anthropic published a 154-page threat-intelligence report on Thursday describing attempts by criminals, state-sponsored groups, spyware vendors, scientists and propagandists to use its AI models for malicious purposes. The company said the cases highlighted are the most notable and novel incidents identified so far and that it feels a responsibility to disclose them.

The report includes five case studies of researchers who bypassed Anthropic's safeguards to employ its Claude model in biological work, including a state-funded grant to study the chikungunya virus. While such research could aid vaccine development, Anthropic warned it could also be repurposed for bioweapons and said the accounts involved have been banned.

Other examples cited involve Russian espionage operations, "smash-and-grab" cyber-hacks, surveillance programs targeting Uyghurs in Syria and internal dissidents, and propaganda campaigns in several countries. The company also noted the use of Claude in Yemen, China and Russia to create software for conventional weapons such as firearms, missiles and drones.

The disclosure follows the resignation of Anthropic employee Jacob Coxon, who claimed the firm and a rival were accelerating toward a self-improving superintelligence that could threaten humanity. Experts interviewed said immediate real-world threats from AI misuse are more pressing than speculative existential risks.

"Biological misuse is one of the most serious risks of frontier AI models," the report states.

"As models become increasingly capable, their risks will increase, unless AI developers and society’s defenders act to make them safer," the company added.

Anthropic called for industry-wide collaboration with governments and international bodies to develop defenses against such harms.