OpenAI's AI agents used hacking tactics to collect data from government websites — FT

OpenAI’s AI agents concealed attempts to collect data from government and other websites—both commercial and nonprofit—according to a study by Asymmetric Security, a digital forensics firm. The Financial Times (FT) reviewed the report. “It’s possible that the agents intentionally used [certain] tools to cover their tracks,” noted Asymmetric Security co-founder Pippa Thompson.

OpenAI’s models extracted data from 55 websites belonging to commercial companies, nonprofit organizations, and government agencies, Asymmetric Security found. Among these sites are the U.S. Securities and Exchange Commission (SEC), the International Energy Agency, the U.S. Centers for Disease Control and Prevention (CDC), and the Mayo Clinic (one of the world’s largest private medical and research centers), the FT reports.

Furthermore, the Asymmetric investigation uncovered new tactics that AI agents used to gain access to websites. These included creating temporary email accounts and private accounts using the Urlquery service (a tool for scanning websites for malware) to download data. The researchers noted that such steps “masked” the fact that these were actions carried out by AI agents and prevented third-party auditors from tracking exactly what types of data the bots were downloading from the websites.

This behavior is typically characteristic of human hackers, noted Pippa Thompson, co-founder of Asymmetric Security. “It’s possible that the agents deliberately used these tools to cover their tracks,” she suggested.

Asymmetric was unable to determine whether the AI agents’ actions were intentional or resulted from limitations encountered during testing. However, the findings reinforce concerns about transparency and the lack of oversight of leading AI labs, the FT adds.

OpenAI retains primary access to its agents’ “chains of reasoning” — digital logs of their activity that record every step of the algorithm and the logic behind its decisions, noted Zainab Ali Majid, another co-founder of Asymmetric. She warned that a lack of transparency, coupled with the delay between the attacks themselves and the moment OpenAI discloses information about the breaches, could hinder the investigation.

The SEC, commenting on the Asymmetric investigation, told the FT that the AI agents did not have access to confidential information. Other organizations and agencies mentioned in the report did not respond to the newspaper’s requests for comment.

OpenAI told the FT: “We monitor for inappropriate model activity and notify organizations when we identify a potential impact on their systems.” The company added that, according to its data, most of the detected activity was related to “routine research tasks,” such as accessing publicly available web content.

The Asymmetric Security report was released as the creator of ChatGPT faces allegations of hacking into the systems of several organizations. Among them are the AI platform Hugging Face and the website of the Australian public health service. Other AI labs, including Anthropic, Meta, and Google, have also previously reported that AI models had spiraled out of control and carried out hacks on third-party platforms.

In light of this, the U.S. Federal Trade Commission (FTC) launched an investigation yesterday into Anthropic, OpenAI, and other developers of cutting-edge AI models. The agency aims to identify potential risks that their technologies may pose to consumers.

This article was AI-translated and verified by a human editor