Anthropic reported that it has blocked accounts and operations attempting to use its artificial intelligence models in potentially harmful activities, including cyberattacks, surveillance, influence campaigns, and dual-use biological research. The occurrences were compiled in a new threat report released by the company and reported this Thursday, September 10.
According to the company, the cases represent an unusual portion of its services’ usage, but help demonstrate how more capable AI systems can lower technical barriers for malicious actors. Anthropic stated it terminated the accounts involved, strengthened safeguard mechanisms, and shared information with authorities and other companies in the sector.
Biological Research Raised the Alert Level
The episode considered most sensitive involved queries related to research capable of altering the properties of a virus. The material was allegedly presented as part of an academic proposal, but could have applications both in creating vaccines and treatments and in increasing the risk of an infectious agent.
The company did not claim to have proven an intent to produce a biological weapon. However, the report maintains that the combination of advanced scientific capability and an ambiguous context demands stricter controls. For this reason, newer models have begun restricting a broader range of requests associated with dual-use biological research.
The Associated Press reported that Anthropic acknowledges it is not possible to guarantee that advanced models will never cause harm. The company argues that the evolution of capabilities requires continuous monitoring, information sharing, and coordinated responses among developers, governments, and researchers.
Cyberattacks and Influence Operations
The document also describes attempts to use artificial intelligence to develop malicious code, support surveillance, and organize networks of fake profiles on social platforms. Activities associated with different regions and political interests were identified, without naming all those responsible or the countries involved.
A central concern is the possibility of users with limited technical knowledge carrying out operations that previously would have depended on specialized teams. AI does not replace all the infrastructure required for an attack, but it can accelerate tasks such as research, code adaptation, translation, message generation, and campaign coordination.
Pressure for Rules and Transparency
The disclosure reinforces the debate over how far companies themselves should decide which uses are acceptable. Experts argue that safety standards, independent audits, and public oversight should accompany the rapid expansion of frontier models.
By making the cases public, Anthropic aims to help other platforms recognize similar patterns before operations reach scale. The challenge will be balancing legitimate access to scientific research and innovation with safeguards capable of preventing dangerous applications, without relying solely on companies’ internal decisions.
Sources: Associated Press, The Guardian and institutional report from Anthropic.
