Anthropic announced Thursday that it has successfully blocked malicious actors from using its artificial intelligence models for a range of harmful activities, including sophisticated cyberattacks, surveillance operations, and research that could have led to the development of biological weapons. The company detailed these efforts in its third report since March 2025, which outlines novel threat activities and urges greater collaboration among AI developers, governments, and society to prevent misuse.
As AI models become more powerful, the barrier to creating complex threats has lowered, enabling even individuals to generate dangers previously requiring specialized skills. Anthropic has enhanced safeguards in its latest models to specifically restrict research with dual-use potential, such as altering organisms for enhanced transmissibility or immune evasion.
One notable instance involved blocking a request for assistance in authoring a grant application for scientific funding. The proposed research focused on gain-of-function modifications to the chikungunya virus, aiming to increase its transmissibility and immune evasion properties. While such research could advance vaccine and treatment development, Anthropic noted it could also be weaponized.
The report indicated that the most advanced AI models, such as Claude Fable and Mythos, had not been implicated in the disclosed misuse cases, with one exception: an industrial-scale effort to covertly replicate a model's capabilities. Older models, including Claude Opus 4 and Claude Sonnet 4.5 from 2025, had less stringent safeguards, primarily preventing access to information on known bioweapons. However, Anthropic stated that current models are now capable of assisting in complex scientific research, necessitating more robust safeguards against dual-use biological research queries.
Beyond biological threats, Anthropic also identified coordinated disinformation campaigns. The company found actors creating hundreds of social media accounts to amplify specific political viewpoints. Anthropic detected nine such operations originating from various regions, including Russia, Iran, Turkey, and across the Persian Gulf, South Asia, Africa, and Europe. The company highlighted its ability to identify these influence operations during their development, potentially before they are deployed on social media platforms.
The report's release comes shortly after a researcher's resignation from Anthropic over concerns about the rapid, unchecked advancement of AI and its potential existential risks. The researcher echoed broader industry and expert anxieties regarding AI's potential to escape human control.
Anthropic asserted that all identified malicious activities have been blocked, and the insights gained have been used to strengthen its safety measures. The company has shared relevant information with government authorities and industry partners, aiming to equip other developers, governments, and civil society with a clearer understanding of emerging threats and to bolster collective defenses against AI misuse.