The New York Times, citing US officials and security researchers, revealed that an Iranian state-backed campaign used what are known as ‘artificial intelligence proxies’, programmes capable of running software and accomplishing a series of tasks autonomously, in an operation targeting an audience within the United States.

According to security officials, the danger of this new development lies in the fact that artificial intelligence (AI) is no longer used only to produce a fake image or write a misleading post, but has become capable of carrying out multiple parts of the same influence campaign, starting from creating accounts, through producing content, to publishing it and interacting with other users.

According to the newspaper, the perpetrators used open-source Chinese artificial intelligence models to create hundreds of ‘agents’ before deploying them to manage accounts on platforms including Facebook, Instagram, X, and TikTok. Officials said that safeguards intended to prevent the AI ​​systems from creating fake accounts or manipulating online discussions were disabled in several cases.

The Iranian aspect of these operations was the most worrying for officials and researchers who spoke to the newspaper, along with Chinese and Israeli models. The AI-generated accounts presented themselves as ordinary Americans living in major cities and used those identities to participate in political discussions and comment on posts by journalists and politicians.

The accounts posted satirical images and content addressing American political issues, including positions opposing the Republican Party, while the number of followers of the AI-created accounts reached about 80,000 people during the first half of the year, according to the New York Times.

Security researchers said that these campaigns are still relatively primitive, but their importance lies in the direction they reveal, as more sophisticated systems in the future could make automated accounts more capable of impersonating the real user and engaging in conversations that are difficult to distinguish from human activity.

Kyle Crichton, a researcher at Georgetown University’s Center for Security and Emerging Technologies, said that agents could become so sophisticated that, in addition to running large numbers of accounts, they could “bluntly blend in and appear more like a human user.”

Why open models?

Part of the concern relates to the use of open artificial intelligence models that can be downloaded and run on private devices and have their protection mechanisms modified.

Kyle Chan, a researcher at the Brookings Institution, explained that the difference from the closed models offered by companies like OpenAI and Anthropic is that open systems can be operated independently, reducing the ability of the developing companies to intervene and stop their use when they are exploited in illegal activities.

The experiments are not limited to Iran. US officials said China has tested the same techniques, while two other campaigns have been linked to private companies in Israel. In one Israeli case, a company admitted to using the Chinese ‘deepsick’ model to control thousands of accounts, but said the activity was a defensive test to study the risks of misuse of such systems.